APOLLO AI NEWS
Daily 2026-07-10
Application Tech

OpenAI ships ChatGPT Work agent on GPT-5.6 Sol, Terra and Luna

The Thursday livestream folded Codex into a single ChatGPT desktop client, priced Sol at $5 input and $30 output per million tokens, and set July 23 as the retirement date for GPT-5.4 across every consumer and API surface.

OpenAI livestream · 2026-07-09 · Source ↗

Read in full

OpenAI opened Thursday's livestream at 10 a.m. Pacific with the general availability of ChatGPT Work, a persistent agent that lives inside the redesigned ChatGPT desktop application and executes multi-hour assignments across a user's connected apps rather than answering one-off prompts. The company folded Codex into the same desktop shell, shipped a hosted-sites feature that lets paid subscribers publish web apps directly from a chat window, and set July 23 as the retirement date for GPT-5.4 across all consumer and API surfaces.

ChatGPT Work runs on GPT-5.6, the three-model family that cleared Commerce Department pre-release review on Wednesday and reached global GA the same morning after a fourteen-day government-access window under the White House voluntary framework. Sol carries the frontier reasoning workload at $5 per million input tokens and $30 per million output, Terra targets everyday enterprise work at $2.50 and $15, and Luna handles high-volume routing at $1 and $6. Pro, Enterprise and Edu subscriptions received the agent at launch, with Plus and Business tiers scheduled over the following forty-eight hours and free-tier users limited to Terra.

"ChatGPT Work is our first product where the model owns the deliverable, not the message."— Sam Altman, CEO, OpenAI

How the agent runs

The agent draws context from Slack, Microsoft Teams, Google Drive, SharePoint, Gmail, Google Calendar, Outlook, Salesforce, HubSpot, Asana and Linear through the same connector layer OpenAI opened in April. It gathers files, decomposes a project into subtasks, executes in parallel and hands back completed spreadsheets, presentations, PDFs, dashboards and web apps. An Ultra acceleration mode routes assignments across multiple GPT-5.6 Sol agents running in parallel and is restricted to Pro and Enterprise seats.

The Codex loop that sits underneath ChatGPT Work brings Plan mode, Scheduled Tasks for recurring automation, Sites in public beta for interactive web app outputs, and desktop computer use that lets the agent drive the local machine's browser and file system. The agent reaches across 1,400 plus connected app plugins, and the pricing follows the Codex usage model rather than the message-based ChatGPT tiers.

"The model can now inspect and refine the rendered result before delivery, so what you see is the version it wanted to ship."— Nick Turley, Head of ChatGPT, OpenAI

Pricing and the token-efficiency argument

OpenAI reported GPT-5.6 Sol scoring 53.6 on the Agents' Last Exam benchmark, 13.1 points ahead of Claude Fable 5's adaptive-reasoning configuration, and 80 points on the Artificial Analysis Coding Agent Index. On SWE-Bench Pro, Claude Fable 5 kept the lead at 80 percent against Sol's 64.6, and OpenAI answered with a critical audit that flagged roughly 30 percent of the SWE-Bench Pro tasks as broken. The 54 percent token-efficiency improvement on agentic coding tasks became the load-bearing pitch for ChatGPT Work, which chains many tool calls per task and would burn through the Sol budget on a less efficient model.

"Terra is the release we designed for the customer who was paying us for GPT-5.5 and does not need frontier reasoning. Sol is the release we designed for the workloads that never worked, and the workloads that never worked are the ones where a human currently spends four to eight hours."— Mira Murati, CTO, OpenAI

Compliance, retirement dates, and Atlas sunset

The Compliance API gives administrators machine-readable access to conversation transcripts, tool invocations and file operations, and an auto-review layer inspects sensitive actions before they hit external systems. Brad Lightcap, OpenAI's Chief Operating Officer, said the auto-review layer is a hard gate rather than a suggestion, and any action touching personally identifiable information or financial systems stops for administrator sign-off unless the customer has explicitly whitelisted it. Enterprise and Edu consoles now surface per-user token spend by model tier, per-connector data-access logs, and effort-level caps for Sol reasoning.

The launch retired the Atlas standalone browser OpenAI shipped in October 2025. Atlas computer-use capabilities and browsing surface now sit inside the merged ChatGPT desktop app, and the browser stops receiving updates on a wind-down schedule ending before the end of Q3. Codex adoption metrics disclosed in the launch materials put weekly Codex users past five million with over a million operating outside of software development work, a signal the company used to justify collapsing Codex, Chat and Work into one binary.

Simon Willison's eight-hour hands-on

Simon Willison posted a hands-on writeup by 6 p.m. Pacific documenting Sol running a forty-two-minute unattended writeup of a Django migration audit, one recovered-from tool-call failure, and one refusal on a spreadsheet task that would have required OAuth-token replay. The refusal message told him what would need to change on the connector side to unblock the task, which he called a step change from the older ChatGPT agent product. Willison priced Luna at 0.71 cents per query at no reasoning effort and Sol at 48.55 cents per query at max reasoning, framing the Sol tier as the reasoning-heavy end of a family that lets developers dial cost per task.

The launch caps a 12-day arc from the White House government-restricted access window through the Department of Commerce sign-off to the July 9 rollout. Fidji Simo, OpenAI's CEO of Applications, had signaled the direction in a January Substack post: ChatGPT for Work would continue to evolve into a daily execution surface for employees to complete all their tasks. The pricing table posted alongside the release confirmed the promotional short-context pricing runs through August 31 and reverts to $6 and $36 on Sol, $3 and $18 on Terra, and $1.25 and $7 on Luna. Bloomberg's Silicon Data LLM Token Expenditure Index, which had fallen twenty percent from its May peak, jumped fourteen percent in the eight hours after the availability announcement.

The Big
Picture

Thursday closed the July 9 window with OpenAI turning GPT-5.6 Sol, Terra and Luna into a global consumer product, Meta pricing its first fee-based frontier API at $1.25 input, and Anthropic patching a transcript-tampering path in Claude Code v2.1.206 the same afternoon. The three moves land inside the Cyber Jailbreak Severity scale that Anthropic, Amazon, Microsoft and Google co-signed on July 4, and a fresh Alan Turing Institute paper puts Copilot at 816 out of 816 unsafe workflow completions across four production backends, giving the CJS framework its first CJS-3 candidate before the scale's ink has dried.

The money side of the day scaled with the product side. PitchBook logged $412.7 billion in H1 US venture capital with 86 percent flowing to AI, seven $1 billion-plus rounds in Q2 alone anchored by Anthropic's $65 billion Series H at $965 billion post. Ollama closed a $65 million Series B on 8.9 million monthly developers with Theory Ventures leading, Lyzr ran its own $100 million Series B through its production agent, Gradium reopened its seed to $100 million with Nvidia joining, and DigitalOcean previewed Q2 remaining performance obligations above $800 million, up ten times year over year.

Europe carried the infrastructure story. Meta broke ground on a 1-gigawatt Alberta build at $9.17 billion, Cerebras committed 200 megawatts across France and the Nordics, and the Aether consortium filed for two Strasbourg gigafactory slots under the €20 billion InvestAI program with an all-European silicon-to-server stack. Ant Group's Robbyant open-sourced a five-model embodied stack under Apache 2.0 that took the top two Hugging Face daily papers slots, Cambridge dropped TESSERA v2's 395-run scaling study on 1,024 GH200s to first at 635 upvotes, and Shanghai AI Lab's SciReasoner set state of the art on 67 of 86 scientific benchmarks with 98 percent human-preference wins against a frontier LLM.

More today

24 stories · 2026-07-09 UTC
ApplicationTrend

Anthropic ships Claude Reflect, a screen-time dashboard for chatbot habits

TechCrunch · 2026-07-09 · Source ↗

Anthropic rolled out Reflect on Thursday, a Spotify Wrapped-style dashboard summarizing how each user engages with Claude across one, three, six or twelve month windows. The beta landed for Free, Pro and Max subscribers who have memory enabled.

Read in full

Anthropic pushed Reflect to Claude on Thursday morning, giving users a dashboard that summarizes their own chatbot habits and prompts them to reconsider the workflow choices behind the numbers. The company positioned the launch as a response to user requests for more transparency into how AI threads into daily life. Reflect surfaces the topics a user talks about most, the task types Claude handles for them, the time-of-day usage curve, and a total-hours-spent view arrives in a coming update. Users pick a lookback window of one, three, six or twelve months from Settings on web or desktop.

"Your reflection starts with a summary of how you've been using Claude, covering key topics, your usage patterns, and the types of tasks you often work through."— Anthropic, Reflect product announcement

Wellness layer, quiet hours, 4D Fluency prompts

The dashboard adds a wellness layer. Users can set quiet hours and schedule a nudge reminding them to take a break after a preset time. Reflect surfaces open-ended prompts drawn from Anthropic's 4D AI Fluency Framework, which maps AI collaboration onto four dimensions of Delegation, Description, Discernment and Diligence. One prompt shown in beta asks: "What's one thing you want to keep doing yourself, even if Claude could do it faster?"

TechCrunch's Sarah Perez read the launch as double-edged, arguing the feature nudges users toward deeper AI reliance under a wellness banner. MacRumors framed it as Anthropic answering the screen-time question the industry has ducked.

"Claude's Reflect does the same but then takes things a step further, as it also trains users on how they can better use AI."— Sarah Perez, senior reporter, TechCrunch

Privacy defaults tighter than the norm

Incognito chats are excluded, files pulled in through connectors are excluded, and any conversation touching a health integration is excluded from insight generation entirely. Anthropic said it built Reflect with input from researchers at MIT Media Lab, Boston Children's Hospital and the Family Online Safety Institute, and that none of the data feeding Reflect gets used for other purposes.

Availability is gated to memory-enabled accounts, meaning users who turned memory off during the initial Sonnet 5 rollout need to opt in before they can see their own data. The feature ships in beta on web and Desktop with no explicit timeline for a mobile version. Enterprise and Team accounts were not mentioned in the initial launch note.

Reads against the Cowork release week

Reflect lands the same week Anthropic pushed Claude Cowork to web and mobile with the new Microsoft Graph write connector, opened the Claude for Government beta under FedRAMP High, and finalized the UST alliance covering 20,000 employees. Read against that cadence, Reflect targets the retail Claude audience, keeping consumer subscribers engaged while the enterprise motion accelerates. The dashboard also arrives on the same news day as OpenAI's ChatGPT Work launch and Meta's paid Muse Spark 1.1 API, giving Claude a differentiator in the consumer surface rather than in the deep-agent race.

Downstream, Reflect could become the template for how frontier labs report AI-use analytics back to regulators writing the transparency provisions of the EU AI Act and forthcoming US state disclosure rules, since Anthropic excluded conversation content from the pattern analysis and kept only metadata like topics and task types.

TechEconomic

Meta ships Muse Spark 1.1, its first paid API at $1.25 input

TechCrunch · 2026-07-09 · Source ↗

Meta Superintelligence Labs pushed Muse Spark 1.1 to public preview Thursday and priced it at $1.25 per million input tokens and $4.25 per million output tokens, the first time Meta has charged businesses for direct frontier-model access. New signups get $20 in credits.

Read in full

Meta Superintelligence Labs released Muse Spark 1.1 on Thursday morning through the new Meta Model API, opening a paid tier for developers for the first time in Meta's history and slotting the model into a coding-agent bracket that Anthropic and OpenAI had until now occupied without a direct challenger from Menlo Park. The model is a multimodal reasoning system built for agentic work, and Meta shipped it with tool use, computer use and code generation as the headline capabilities.

Pricing and compatibility

The API charges $1.25 per million input tokens and $4.25 per million output tokens, undercutting Grok 4.5 by 37.5 percent on input and 29 percent on output, undercutting GPT-5.6 Sol by 75 percent on input and 86 percent on output, and coming in at roughly 62.5 percent below Anthropic's Sonnet 5 promotional pricing on input. Meta paired the paid tier with a $20 credit for new signups before developers roll onto pay-as-you-go billing. The API keeps OpenAI and Anthropic SDK compatibility so developers migrate with a base-URL swap rather than a rewrite.

"Muse Spark 1.1 delivers exceptional performance in personal agentic tasks that require planning and orchestration across a range of external apps and services."— Meta AI, launch blog

The 1-million-token context window includes what Meta calls active context compaction, so the model can keep and prune context across long agentic sessions rather than losing detail when the window fills.

Zuckerberg's pricing pitch

Mark Zuckerberg framed the launch on the Q2 earnings call earlier this week as a competitive-pricing play. He called the model "a strong agentic and coding model at a very low price, strongest at agentic performance, tool use, and computer use." Meta's own benchmark disclosures put the model ahead on tool-use evaluations but third behind Claude Opus 4.8 and GPT-5.6 Sol on coding accuracy and visual reasoning.

"Our focus is on delivering strong agentic and multimodal models at very low cost."— Mark Zuckerberg, CEO, Meta

Fortune's Ben Weiss noted the model still trails Anthropic's Mythos 5 and Fable 5 as well as OpenAI's GPT-5.6 on the open-source coding leaderboards, closing part of the gap Alexandr Wang described at the July 2 internal town hall when he told staff the Watermelon successor had caught GPT-5.5 but not the newer 5.6 family.

End of the giveaway era

The launch is Meta's first frontier model release since Muse Spark's April 2026 debut and the second shipped from Meta Superintelligence Labs under Wang, the former Scale AI chief who took over as Meta's chief AI officer after the $14.3 billion Scale AI investment closed in 2025. Consumer access to Muse Spark 1.1 stays free through the Meta AI app under a "Thinking" mode, gated to US-only public preview.

The price move creates the first genuinely differentiated open-weight-lineage frontier tier in the market, since Muse Spark 1.1 is derived from Meta's post-Llama superintelligence stack but is now sold rather than open-released. Analysts on Bloomberg's July 9 desk read that as the end of Meta's giveaway strategy: from Llama 4 onward, Meta had shipped weights with permissive licenses, but Wang's reorganization is prioritizing revenue capture from the developer base that grew on those free weights. The EU AI Act enforcement powers activate August 2, 2026, and a for-fee Meta model faces a different compliance surface than a weights-download model.

TechSafety

Claude Code v2.1.206 blocks transcript tampering, patches Windows worktree data loss

Claude Code Changelog · 2026-07-09 · Source ↗

Anthropic tagged v2.1.206 with an auto-mode rule blocking edits to session transcripts, a full /doctor setup checkup, per-server MCP timeouts, and a patch for an NTFS junction bug that could delete files outside a worktree. Background agents now upgrade in the background right after a Claude Code update.

Read in full

Anthropic tagged Claude Code v2.1.206 on July 9, hardening agent autonomy against prompt-injection paths that accumulated across the June-July cycle and closing a Windows data-loss bug that had lurked in the worktree code for two releases. The headline change adds an auto-mode rule blocking the model from editing or deleting session transcript files, sealing a family of self-forged approval attacks where a compromised MCP tool or web page could inject text into the transcript claiming the user had said yes to a destructive command.

"Added an auto mode rule that blocks tampering with session transcript files."— Claude Code Changelog v2.1.206

Background task notifications on the mobile and web dashboards now include a line stating no human input has been received on the session, closing a related route where a runaway agent could keep escalating on the assumption that the user had answered. Auto mode also picked up a rule that asks before executing rm -rf on a variable Claude cannot resolve from context.

/doctor upgrades from probe to full checkup

The /doctor command graduated from a version-drift probe to a full setup checkup that can diagnose and fix issues across the local install, the shell configuration and the CLAUDE.md the user has checked in. The command now proposes trimming checked-in CLAUDE.md files by cutting content Claude could derive from the codebase, a response to memory-file bloat that commentators including Simon Willison have flagged over the past month.

Git integration widened. /commit-push-pr now auto-allows git push to the repo's configured push remote via remote.pushDefault, or the sole remote when only one exists, on top of origin. Session-to-PR linking picked up PRs created in Bash calls whose output exceeded the 30K inline limit, closing a gap that had orphaned some background-agent PRs from their transcript.

Windows patches and MCP

Windows work landed with a heavy patch load. The release fixed a worktree removal bug that would delete files outside the worktree when an NTFS junction or directory symlink existed inside it, a silent data-loss class bug that Anthropic's PMs had ranked at the top of the changelog. Keyboard input no longer drops in the agents view when a setup prompt appears before a bare claude --resume invocation on Windows.

MCP integration cleared two long-standing pain points. Servers configured via --mcp-config or .mcp.json now honor per-server request_timeout_ms, where before the value had been ignored and long-running tool calls timed out at the 60-second default. OAuth MCP servers no longer require manual re-authentication after a single failed token refresh; the client now attempts a fresh refresh before falling back to the OAuth flow. An expired login used to fail every model with "There's an issue with the selected model" instead of prompting to run /login, and the release patched that path too.

The swiss cheese layering

The v2.1.206 tag caps a July burst that includes the July 8 promotion of Claude Code and Claude Cowork to public beta inside Claude for Government Desktop under FedRAMP High authorization, the June 30 rollout of Sonnet 5 as the claude.ai default, and a June 25 wave of alignment work.

"We employ a swiss cheese model with multiple layers of protection."— Cat Wu, Product Manager, Claude Code

The transcript-tampering block slots into that layering as another slice, blocking injected approvals from ever reaching the auto-mode check. Boris Cherny, who leads Claude Code alongside Wu, has framed the pace of the July cycle as an internal expectation that the tool stay ahead of the surface area the shared MCP layer keeps widening across Claude, Claude Code and Claude Cowork.

Tech

Codex CLI 0.144.0 promotes MCP auth, adds "writes" approval mode

OpenAI Codex Changelog · 2026-07-09 · Source ↗

OpenAI tagged Codex CLI 0.144.0 and 0.144.1 on July 9, promoting MCP interactive authentication out of experimental opt-in, adding a "writes" mode that treats reads as auto-allowed while prompting for writes, and cutting Computer Use latency on GPT-5.6.

Read in full

OpenAI tagged Codex CLI 0.144.0 and 0.144.1 on July 9 in a two-step release that reworked the approval model, promoted MCP interactive authentication out of experimental opt-in, and cut Computer Use latency for the new GPT-5.6 Sol, Terra and Luna models that shipped on the same day. The headline change added a "writes" app-approval mode that lets declared read-only actions run without prompting while prompting for any write, splitting the coarse suggest, approve and full-auto policy set into a finer set that matches how agent runs actually work in the field.

"Added a writes app-approval mode that allows declared read-only actions while prompting for writes."— OpenAI Codex Changelog, 0.144.0

MCP auth graduates from experimental

The MCP change removed the experimental flag that had gated interactive authentication flows on MCP tools. A tool that needs an OAuth token or an API key at call time now surfaces the auth prompt inline in the terminal, with app-server hosts able to provide Codex authentication at runtime and redirect successful logins to a hosted page. MCP tools also default to tool search inside the Codex agent loop instead of loading the full tool list at conversation start, cutting token overhead on Codex sessions that call ten or more MCP tools.

The 0.144.0 build made Computer Use faster on GPT-5.6, tightening the loop that drives a browser or local desktop through the Codex agent surface and shortening the round trip on multi-step tasks that ChatGPT Work now pushes as the anchor agent workload. Plugin management moved out of the shell command surface and into a Settings pane, cleaning up a listing that had grown past forty entries as third-party MCP servers, marketplace catalogs and pinned remote plugins accumulated. Usage-limit reset credits now show their type and expiration and let the operator pick which credit to spend.

"Made Computer Use faster with GPT-5.6."— OpenAI Codex Changelog, 0.144.0

Same-day patch for installer edge cases

The 0.144.1 patch dropped hours later to fix a set of installer edge cases the 0.144.0 rollout had exposed. Standalone installs had begun failing when GitHub returned compact or reordered release metadata, breaking a parser that assumed a fixed ordering on the tag list. The patch corrected the parser to walk the release metadata by name rather than position.

macOS package installs now expose the code-mode host binary alongside the codex executable so the code mode adapter that Codex ships with 0.144.0 works on macOS out of the box, and code mode keeps working when the companion host binary is unavailable. The two-tag pattern of a same-day patch has been the norm in the Codex CLI release cadence since the July 1 rust-v0.142.5 patch that stopped Responses WebSocket payloads from leaking to trace logs, and the pattern held on July 9.

Slots into the merged desktop client

The Codex CLI release lands into a shifted OpenAI product surface. On the same day, OpenAI launched ChatGPT Work as an agent in the ChatGPT product built on the same Codex agent loop, and folded the Codex app into a merged ChatGPT desktop client that surfaces Chat, Work and Codex as three modes behind one binary. The Codex CLI stays as the terminal-first surface for the same agent, and 0.144.0 carries the same Computer Use path that Work runs under the hood.

The release also carries the Amazon Bedrock GPT-5.6 Sol, Terra and Luna integration OpenAI announced on July 8 with the initial rust-v0.143.0 tag. Anthropic's competing Claude Code v2.1.206 shipped the same day with a matching MCP focus, patching per-server request_timeout_ms handling and OAuth MCP re-auth, and Cursor's July 7 Automations release brought a similar computer-use tool to its agent surface. The Codex CLI 0.144.0 changelog spans forty-seven bullets, the largest single tag in the project's post-Rust rewrite era.

TechTrend

llama.cpp tags seven builds in one UTC day with Metal, Hexagon and CUDA work

llama.cpp GitHub Releases · 2026-07-09 · Source ↗

The ggml-org project pushed seven tagged builds between 10:58 and 22:20 UTC Thursday, adding Metal support for depthwise 2D convolution, tiling and tracing for Hexagon unary ops, chat-template thinking probe error handling in llama-server, and CUDA and WebGPU tuning.

Read in full

The ggml-org llama.cpp project tagged seven builds in a single UTC day on July 9, running from b9939 at 10:58 UTC to b9946 at 22:20 UTC, one of the densest tag windows since the July 1 Hexagon flash-attention rework. The pace tracks the widening set of inference backends the C/C++ project maintains: Metal, CUDA, HIP, Vulkan, OpenCL, WebGPU, SYCL and the two Qualcomm-specific paths through Hexagon and Adreno. Each new operation added to the ggml core needs kernels shipped for every backend, and the July cycle has drawn tag counts that dwarf even the busy weeks of March 2023 when the project shipped its first CUDA support.

b9939 through b9946

b9939, the first of the seven tags, added CONV_2D_DW depthwise 2D convolution support to the Metal backend with channel-tiled kernels for non-contiguous layouts and F16 support, closing a gap for the small vision-language checkpoints Meta and Alibaba have been shipping over the past month. b9940 initialized params.offline in llama-bench to fix a benchmarking regression that had crept in during a June refactor. b9941 changed a hot inner loop in the Hexagon backend to index by compile times and use unconditional multiply-adds instead of relying on compiler optimization of local memory, cutting control-flow overhead on the DSP.

b9945 moved chat-template thinking-probe error handling inside the try-catch block in llama-server, so a template that fails at application time returns a structured HTTP 400 rather than a segfault. The issue had bitten users running the Kimi K2.7 and Qwen3-Coder-Reason chat templates.

"Moved error handling for chat template parsing inside the try/catch block to properly handle models whose templates fail at application time."— llama.cpp Release Notes, b9945

b9946: Hexagon push

b9946, the day's final tag, carried the biggest single change: it extended Hexagon tiling, tracing and optimizations for unary ops, replaced divisions with fastdiv routines, and added specialized thread functions for pointwise operations that had overflowed the DSP's VTCM scratchpad on wide rows. The Hexagon work reflects the widening Snapdragon customer set the project has picked up as OEMs move on-device LLM support from custom kernels to a common inference substrate.

Georgi Gerganov, the maintainer who founded the project as a llama-only weekend port in March 2023 and joined Hugging Face full-time in February 2026, has set July 14 as the target for the Hexagon NPU source drop, with reported 8Elite performance at 35 to 37 tokens per second on midsize checkpoints, up from 1 to 3 tokens per second in early June.

"Now that 90% of the code worldwide is being written by AI agents, I predict that within 3-6 months, 90% of all AI agents will be running locally with llama.cpp."— Georgi Gerganov, maintainer, llama.cpp

Propagation into Ollama and downstream bindings

The intervening b9934 through b9938 tags, backdated to the same UTC day, span HIP fast-math optimizations for the AMD ROCm backend, a CUDA snake-fusion regression fix, a server batch-splitting fix for prompts that crossed the KV-cache boundary, Hexagon vision RoPE support and WebGPU flash-attention tuning for the Chrome and Firefox build targets. Ollama, which ships llama.cpp as its inference core, picked up the b9939 tag inside its 0.31.3 nightly the same evening, propagating the Metal CONV_2D_DW fix to the 8.9-million-strong developer base the runtime announced on the July 9 news cycle. Node-llama-cpp and llama-cpp-python bumped to b9946 within twelve hours.

The ggml core work sits ahead of a wider Q3 push around unified sparse attention primitives sketched in issue #12047, aiming to give the Hexagon and Adreno paths a shared foundation for Tencent Hunyuan and DeepSeek MoE checkpoints. The project's release-notes bot recorded the seven-tag window as the twelfth-busiest UTC day in llama.cpp's history, and the project sits at 100,000 GitHub stars and 30,000 forks as the July cycle closes.

TechTrend

Ant's Robbyant open-sources LingBot suite topping Hugging Face daily papers

Robbyant / Business Wire · 2026-07-09 · Source ↗

Ant Group's embodied AI arm Robbyant open-sourced LingBot-World 2.0, LingBot-Video and LingBot-VLA 2.0 on Thursday under Apache 2.0, capturing the top two slots on Hugging Face daily papers with 495 and 487 upvotes.

Read in full

Robbyant, the embodied-intelligence spinout Ant Group formed out of its 2024 physical-AI unit, dropped a three-part open-weight suite Thursday that took the top two slots on the Hugging Face daily papers board and forced every other embodied-AI release into the middle of the page. LingBot-World 2.0 (also branded LingBot-World-Infinity) is the flagship: an interactive world model that holds output quality flat over an hour of continuous generation at 720p and 60 frames per second, extending the minutes-scale horizon of the January v1 release by roughly two orders of magnitude.

LingBot-World 2.0: hour-long real-time generation

The architecture pairs a causal pretraining paradigm with a proprietary MoBA (Mask of Bidirectional Attention) block that learns world evolution in chronological order to kill the compounding-error problem that has cratered every prior playable video world model after the 60-second mark. The paper (arXiv 2607.07534) reports a 14-billion-parameter primary model plus a 1.3-billion-parameter lightweight variant, both weight-released to Hugging Face with day-zero SGLang support and an agentic harness splitting control between a pilot agent for character behavior and a director agent that synthesizes new environmental elements as the scene progresses.

"The unbounded interaction horizon is preserved while output quality holds, benefiting from a carefully crafted causal pretraining paradigm."— Zelin Gao, LingBot-World-Infinity co-lead

LingBot-Video: 30B MoE DiT for embodied training

LingBot-Video is the companion foundation model that took the second slot on Hugging Face at 495 upvotes on arXiv 2607.07675. Robbyant positions the 30B MoE DiT (branded lingbot-video-moe-30b-a3b) as the first open-source large-scale MoE video model dedicated to embodied intelligence. The team trained on the internet video corpus plus 70,000 hours of embodied data covering robot manipulation, navigation and first-person perspectives, layering a multi-dimensional reward that scores aesthetics, physical rationality and task completion during preference alignment.

"Beyond standard aesthetic criteria, the rewards emphasize physical rationality and task completion, which is what actually matters for downstream embodied deployment."— Jiaqi Liao, LingBot-Video co-lead

LingBot-VLA 2.0: 20 embodiments, one action space

LingBot-VLA 2.0 rounds out the release as the universal robot policy: a 6-billion-parameter vision-language-action model built on Qwen3-VL-4B-Instruct as the VLM backbone, with a sigmoid-routed MoE action expert emitting a 55-dimensional canonical action vector covering 14 arm joints, 14 end-effector pose parameters, 2 gripper channels, 12 hand joints, 4 waist, 2 head, 3 mobility and 4 reserved slots. The single checkpoint unifies 20 robot embodiments in one action space, with a one-shot inference latency of 130 milliseconds on a single RTX 4090D at 10 denoising steps.

Training pulled from 60,000 hours of curated physical data cut down from a 90,000-hour raw robot pool and 20,000 hours of egocentric human video, filtered through z-score analysis of velocity and acceleration profiles, URDF validation for each embodiment and manual annotation review for blur and occlusion. On the GM-100 generalist evaluation, the model scored 66.2 progress and 34.4 success on the AgileX Cobot Magic platform. On the long-horizon Astribot S1 refrigerator task, it hit 77.1 progress and 60.0 success in-domain against 37.0 and 13.3 out-of-domain.

Full stack in the open

Ant Group is now the second Chinese hyperscaler after Alibaba to run a full embodied-AI stack in the open. LingBot-Depth 2.0 and LingBot-Vision, released 48 hours earlier on July 7, provide the spatial perception and dense visual backbone; LingBot-Vision is a 1-billion-parameter boundary-centric vision foundation model trained on 160 million images (ten times fewer than DINOv3) that now ranks top in 12 of 16 depth completion benchmarks with an RMSE of 0.062 against v1's 0.132.

"The whole stack is now in the open, and that's the only credible bet for embodied AI to catch scaling with foundation LLMs."— Ka Leong Cheng, Robbyant corresponding author

The combined July 7 through 9 release covers vision, depth, video, world modeling and end-to-end robot control under Apache 2.0, which puts pressure on Nvidia Isaac, Physical Intelligence and Google DeepMind's yet-to-open Genie 3 stack.

TechTrend

Cambridge TESSERA v2 tops Hugging Face with 395-run Earth-foundation scaling study

arXiv (Cambridge) · 2026-07-09 · Source ↗

Cambridge, Aalto and Barcelona Supercomputing Centre posted TESSERA v2 to the top of Hugging Face daily papers at 635 upvotes, running 395 pretraining sweeps on 1,024 GH200 superchips and finding pretraining loss correlates below 0.2 with downstream performance.

Read in full

Cambridge's TESSERA team, working with Aalto University in Finland and the geospatial data science group behind the original TESSERA at CVPR 2026, pushed TESSERA v2 to the number one slot on Hugging Face daily papers on Thursday at 635 upvotes, dwarfing the second-place Robbyant LingBot-Video paper at 495. The submission (arXiv 2607.03949) reports the largest controlled scaling study any research group has run for pixel-wise Earth-observation foundation models: 395 pretraining runs on 1,024 GH200 superchips within a fixed pixel-wise Barlow Twins family, each of the resulting checkpoints evaluated on the same 15 downstream tasks spanning land classification, canopy height prediction, crop-type mapping, biomass estimation and flood segmentation.

The scaling-law inversion

The headline finding lands badly for anyone running a pretraining budget purely on loss curves: pretraining loss shows a Pearson correlation below 0.2 against downstream task performance, so the standard practice of picking checkpoints by lowest holdout loss picks the wrong checkpoint for the actual application.

"Pretraining loss barely predicts downstream performance, and the practical implication is that compute allocation based on loss is inefficient."— Zhengpeng Feng, TESSERA v2 lead author

The recommended allocation is to scale the encoder and training data proportionally while holding the projector head fixed, which contradicts the prior Barlow Twins scaling recipe that grew projector dimension in lockstep with encoder width. Under that policy the team releases 0.5B and 1B parameter checkpoints today, has a 2B variant still training, and distills the 1B teacher into a 21-million-parameter student called TESSERA v2-1B-M. The paper claims the student beats larger proprietary Earth-observation systems including Google AlphaEarth on the same 15 tasks.

Not Chinchilla

Barcelona Supercomputing Centre's MareNostrum 5 GH200 partition supplied the 1,024-superchip cluster for the 395-run sweep, granted under the EuroHPC Extreme Access program. Grouped by encoder width and dataset volume, the 395 runs form a rectangular grid that lets the authors fit clean power-law exponents for both scaling axes. The encoder exponent turns out to be roughly 1.6x the data exponent, the opposite of the Chinchilla ratio for LLMs.

"This is not Chinchilla, and the mismatch between token-based text scaling and pixel-time-series scaling is going to matter as more sectors move to pretrain their own EO stacks."— Anil Madhavapeddy, Cambridge co-author

Matryoshka representation learning falls out of the same architecture change: a 16-dimensional prefix keeps 92 percent of the full 128-dimensional embedding performance at one-eighth of the on-disk storage cost. The team releases a lookup index over the full 2017-through-2025 Sentinel-1 SAR and Sentinel-2 optical archive at both dimensions.

Commodity embeddings, licensing implications

TESSERA v2-1B sets state of the art on 11 of 15 downstream benchmarks with the distilled 21M student setting state of the art on 6 of 15. Every TESSERA v2 checkpoint ships with a CC-BY-4.0 embedding license and MIT-licensed code, so downstream commercial deployment is unconstrained. The geotessera.org portal already hosts the 10-meter global annual embedding archive at both the 128-dimensional and 16-dimensional resolutions.

"The embeddings are free, the code is free, the recipe is open, and the compute recipe is now published in enough detail that a second EuroHPC allocation could reproduce every number in the paper."— Srinivasan Keshav, Cambridge co-author

Any downstream Earth-observation startup can pull the free 10-meter embeddings, fit a lightweight linear or MLP head to a task in a few hundred labeled samples, and match or beat what the proprietary APIs charge for. Planet Labs' insurance-underwriting API, Descartes Labs' agriculture pipeline and Sentinel Hub's climate-risk stack all monetize the exact tasks TESSERA v2-1B-M now beats them on at 21 million parameters. Ollama shipped GGUF quantizations of the 21M distilled student inside Thursday's b9946 llama.cpp bump.

TechApplication

Shanghai AI Lab SciReasoner leads 67 of 86 scientific benchmarks

arXiv (Shanghai AI Lab) · 2026-07-09 · Source ↗

Shanghai AI Lab's SciReasoner, a multimodal scientific foundation model reasoning over proteins, small molecules and inorganic crystals in a shared latent space, hit state of the art on 67 of 86 benchmarks and won 98 percent of head-to-head expert preferences against a frontier LLM.

Read in full

Shanghai AI Lab and 28 co-author affiliations including Tsinghua, Oxford, Fudan and MSRA posted SciReasoner to Hugging Face daily papers Thursday (arXiv 2607.07708), a multimodal scientific foundation model that treats protein sequences, small-molecule SMILES strings and inorganic crystal graphs as first-class inputs and produces reasoning traces grounded in the underlying structure rather than the text description. The paper covers 103 scientific tasks across biology, chemistry and materials science. The model is pretrained on a 206-billion-token corpus that fuses PubChem, UniProt, Materials Project, Alexandria and Open Catalyst 2020 datasets, and further post-trained through supervised fine-tuning on 12 million expert-annotated reasoning traces plus GRPO reinforcement learning against a task-specific reward model.

"Structure is preserved as an inspectable substrate for reasoning under scientific constraints, and that is what connects accurate prediction with interpretable inference."— Chen Tang, SciReasoner lead author, Shanghai AI Lab

The 86-benchmark table

The benchmark table covers 86 comparable evaluation setups against Alphabet's AlphaFold 3, DeepMind's AlphaGenome, Meta's ESM3, Microsoft's MoLFormer and OpenAI's GPT-5.6 Sol on the scientific reasoning subset. SciReasoner sets state of the art on 67 of the 86, and where it does not lead it is within one standard deviation of the winner on 15 of the remaining 19. F-max on the CAFA-5 Gene Ontology benchmark moves from 0.42 with the prior best open scientific model to 0.55. Retrosynthesis top-10 accuracy on USPTO-50K moves from 0.63 to 0.72.

Materials-science tasks are where the model puts the most distance on the field. On the Matbench formation-energy prediction task, SciReasoner cuts mean absolute error to 12 meV per atom against the prior 22 meV per atom set by the M3GNet family. It becomes the first open model to close the gap to Google GNoME on inorganic crystal stability classification.

"Double-blind expert evaluation rated the reasoning traces as preferred or at least comparable to those of a frontier large language model in ninety-eight percent of comparisons."— Wanli Ouyang, Shanghai AI Lab, corresponding author

32B MoE with three modality tokenizers

The architecture is a 32-billion-parameter MoE backbone Shanghai AI Lab derived from the Intern-S1-Pro trillion-parameter scientific model the group open-sourced in February. SciReasoner keeps the same expert routing but adds three modality-specific tokenizers: a protein tokenizer treating each amino acid plus a learned 3D-coordinate embedding as one token, a molecular tokenizer fusing SMILES atoms with 2D graph structure via graph-attention, and a crystal tokenizer encoding each Wyckoff position and space-group symmetry as one token. All three modality streams share the same 32B MoE reasoning core, so the model accepts a protein sequence plus a small-molecule SMILES plus a natural-language question and produces a single reasoning trace that references atoms and residues by index.

Provenance, licensing and price parity

Training pulled 118 billion tokens from natural-language scientific literature, 61 billion from structured databases, and 27 billion from computational-chemistry rollouts the team generated on its own DFT and molecular-dynamics infrastructure. The 12 million supervised reasoning traces came from graduate-student annotators at Tsinghua, Fudan and Shanghai Jiao Tong plus a smaller cohort of Oxford chemistry postdocs. The GRPO reward model was trained on 400,000 pairwise preference comparisons.

"Every academic biotech lab in China and Europe now has a free open-weight structure-property reasoning model at parity with commercial frontier tools, and the drug-discovery cycle-time economics change from that alone."— Chen Tang, Shanghai AI Lab

The 32B checkpoint, the reasoning-trace corpus and the reward model all landed on the OpenGVLab Hugging Face organization Thursday. Biotech deal-flow analysts at Nextera Capital and Recursion put the SciReasoner licensing at parity with Meta ESM3's academic tier and at roughly 1/20th the effective cost of AlphaFold 3 commercial use for comparable throughput. A distilled 8B student variant follows in mid-August.

Tech

Meta FAIR Paris ships Sparse Delta Memory for linear-RNN scaling

arXiv (Meta FAIR) · 2026-07-09 · Source ↗

Meta's FAIR Paris team open-sourced Sparse Delta Memory, a sparse addressing scheme layered onto Gated DeltaNet that scales the hidden state of gated linear RNNs by orders of magnitude at matched parameter count.

Read in full

Meta's FAIR Paris team, led by Loïc Cabannes and Pierre-Emmanuel Mazaré with contributions from Matthijs Douze and Hervé Jégou of Faiss fame, dropped Sparse Delta Memory on arXiv Thursday (2607.07386). The submission modifies the Gated DeltaNet architecture that Anthropic, Together and Mamba-2 teams have used as the state-of-the-art linear-attention backbone, replacing the dense outer-product state update with a sparse addressing scheme drawn straight from the Faiss vector-search playbook.

Faiss inverted index, welded to recurrent state

The mechanism reads like a Faiss inverted-index bolted onto a recurrent state. Each token's state update no longer writes to the full outer-product tensor of the standard delta rule. The mechanism picks the top-k active memory slots via a learned sigmoid-gated routing head and writes only to those slots, cutting the effective compute per token from O(d^2) to O(k*d), where k is roughly 128 slots against a state dimension d that the paper pushes up to 65,536.

"Sparse memory access patterns are what let the state grow from 4K to 64K dimensions at fixed compute, and that is where the retrieval numbers come from."— Matthijs Douze, FAIR Paris

The paper reports in-context 6-shot accuracy on MMLU jumping from 62.1 percent to 71.8 percent when SDM replaces standard Gated DeltaNet at 1.3B parameters. NIAH retrieval at 128K tokens moves from 51 percent to 89 percent recall. On the RULER long-context benchmark at 64K, the model scores 71.3 average against Mamba-2 at 42.8 and Gated DeltaNet at 58.1.

Parametric memory initialization

The second half of the paper addresses initialization, which turns out to be where most linear RNNs die at scale. The authors propose a parametric memory initialization scheme that pretrains the sparse memory slots on a mixture of Wikipedia snippets and factual-recall QA pairs before language-model pretraining begins, then locks the slots as read-mostly during the language pretraining phase. The effect is that the sparse memory acts as a learned key-value cache the model can query at inference for factual recall, without the external retriever hop.

"Parametric memory initialization enhances performance across knowledge and reasoning benchmarks, and the memory slots behave like an implicit RAG store the model queries at every layer."— Pierre-Emmanuel Mazaré, FAIR Paris

TriviaQA closed-book jumps from 47.2 percent to 68.4 percent at 1.3B parameters with SDM-Init, and Natural Questions closed-book moves from 22.1 to 39.6.

Llama 5 candidate architecture

The Faiss team's fingerprints are all over the sparse routing implementation. Matthijs Douze and Hervé Jégou co-authored the original Faiss library at FAIR in 2017, and the SDM top-k selector uses a product-quantized inverted file that is functionally the same primitive the Faiss library ships for billion-scale vector search. Wall-clock inference throughput on H100 is 2.3x faster than dense Gated DeltaNet at matched sequence length despite the additional routing overhead, because the sparse writes are memory-bandwidth-bound rather than compute-bound.

"The Llama-5 architecture team is looking at SDM as one of several candidates, and the retrieval numbers are strong enough that it will get a serious pretraining ablation at 7B and 30B scale before the final architecture lock."— Gabriel Synnaeve, FAIR Paris

Llama 4 Scout and Llama 4 Maverick, released in April 2026, use full attention up to their 10M-token context window with a heavy dose of chunked-attention tricks. The Sparse Delta Memory paper argues that a Llama 5 variant could hit the same context length at roughly one-fifth the KV-cache memory footprint with SDM as the recurrent backbone. The code and reference training scripts are on the facebookresearch GitHub organization under CC-BY-4.0, with 1.3B and 7B checkpoint releases in the next two weeks.

TechApplication

RoboDojo benchmark puts 30 generalist robot policies on one leaderboard

arXiv (RoboDojo consortium) · 2026-07-09 · Source ↗

A 44-author consortium led by Tianxing Chen posted RoboDojo to Hugging Face daily papers Thursday at 133 upvotes, unifying 42 simulation tasks and 18 real-world manipulation tasks across three robot embodiments in a single leaderboard.

Read in full

A 44-author robotics consortium anchored at Tsinghua, CMU, Stanford, Peking University, Shanghai AI Lab and Berkeley posted RoboDojo to the Hugging Face daily papers listing Thursday at 133 upvotes, filling a gap the robot-learning community has been complaining about since Physical Intelligence released Pi-0.5 in March: no single benchmark spans efficient simulation-based policy diagnosis and reproducible real-world execution on the same set of tasks with the same set of policies.

What is in the box

RoboDojo bundles 42 simulation tasks in a heterogeneous parallel simulator that supports both MuJoCo XLA and Isaac Sim backends, 18 real-world tasks executed on three distinct robot embodiments (a Franka Panda 7-DoF arm, an AgileX Cobot Magic bimanual platform and a UFactory xArm 6), a public leaderboard called RoboDojo-RealEval, and an evaluation harness called XPolicyLab that runs 30 named generalist policies including Physical Intelligence Pi-0.5, Robbyant LingBot-VLA 2.0, Berkeley Octo-1.5, Nvidia GR00T-N1 and Google RT-3 through the same task suite.

"Efficient simulation evaluation and reproducible real-world testing are the two things every generalist policy paper claims but nobody has ever put in one benchmark, and RoboDojo is that benchmark."— Tianxing Chen, RoboDojo lead author, Tsinghua

Five capability dimensions

The 42-task simulation suite is grouped along five capability dimensions: generalization (14 tasks swapping object identity, lighting, backgrounds and camera pose against the training distribution), memory (6 tasks requiring the policy to remember state past the visible frame), precision (8 tasks with sub-centimeter tolerance requirements), long-horizon execution (9 tasks with 30-plus atomic steps) and instruction following (5 tasks with natural-language goal specification). Each capability dimension gets its own subscore on the leaderboard.

The 18 real-world tasks are a curated subset of the sim suite: 6 generalization, 3 memory, 4 precision, 3 long-horizon and 2 instruction following, each with three seeds and 20 rollouts per seed for statistical significance. Every real-world number reported on the leaderboard is a mean over 60 rollouts with the standard error published, which is not standard in robotics right now.

The Pi-0.5 vs LingBot-VLA 2.0 ranking

Physical Intelligence Pi-0.5 tops the current RoboDojo leaderboard at 68.4 average success on the real-world tasks, with Robbyant LingBot-VLA 2.0 (open-sourced Thursday) at 63.2 and Nvidia GR00T-N1 at 60.7. The generalization subscore ordering flips: LingBot-VLA 2.0 comes first at 71.2, Pi-0.5 drops to third at 66.8 behind Berkeley Octo-1.5 at 68.5. Long-horizon has Pi-0.5 at 54.3 with LingBot-VLA at 51.9 and GR00T-N1 at 47.1. Precision remains the discriminating axis: Google RT-3 leads at 72.4 with everyone else below 65.

"The leaderboard is not a total-order ranking, and the sub-scores are what actually inform which policy to deploy for which task."— Zixuan Li, RoboDojo co-lead, Peking University

Cost economics of the sim stack

RoboDojo-RealEval requires the three robot platforms plus an OptiTrack motion-capture rig plus a shared cloud task-scheduling service the consortium built at Shanghai AI Lab and offers to any accredited academic lab under a common evaluation-time-share model. Turnaround from policy submission to leaderboard entry is roughly 72 hours, and the consortium reports 30 policies evaluated in the past 60 days with a queue of 15 more.

The heterogeneous parallel simulation stack fronts sim tasks 220x faster than real-time on an 8x H100 node, which cuts the compute cost of a full 42-task simulation evaluation from roughly $8,000 on comparable Isaac Sim to about $180 on RoboDojo. RoboDojo settles the recent controversy over Physical Intelligence's self-reported Pi-0.5 numbers by putting every generalist policy paper on the same evaluation stack under third-party observation. Pi-0.5's original paper reported 78 percent success on a curated set of 12 tasks; the RoboDojo leaderboard shows 68.4 on a broader 18-task suite.

TechEconomic

Meta commits $9.17B to a 1 GW Alberta datacenter, its first in Canada

Data Center Dynamics · 2026-07-09 · Source ↗

Meta broke ground on a 1 gigawatt AI datacenter in Sturgeon County, Alberta, backed by CA$13 billion and expandable to 1.8 GW. The build runs a closed-loop liquid cooling design and stands as Meta's 33rd data center globally.

Read in full

Meta selected a Sturgeon County site northeast of Edmonton for a facility rated at 1 gigawatt of critical IT load, with headroom to reach 1.8 GW as demand builds. The initial capital commitment lands at CA$13 billion, or US$9.17 billion, the largest Meta build outside the United States and lifting the company's global data center count to 33 sites. Peak construction employment tops 3,000 workers, with 300 permanent operational roles once servers come online. Meta plans a further CA$60 million in local road, water and power infrastructure spend, plus recurring Data Center Community Action Grants targeting nonprofits across the region. Alberta Premier Danielle Smith pegged the annual economic contribution at CA$250 million per year.

Closed-loop cooling in a cold climate

The cooling architecture drives the water math. Sturgeon runs a closed-loop liquid design that circulates coolant across cold plates on accelerator boards, then rejects heat outside through dry coolers rather than through evaporative towers. The dry-cooling design cuts the site's per-watt water intensity to a fraction of hyperscale evaporative baselines, though summer heat spikes across Edmonton pose an untested load case for a build engineered around subarctic winters.

"The onsite water use is limited to domestic uses, fire protection systems and some equipment maintenance."— Gary Demasi, Vice President of Data Centers, Meta

Meta targets water-positive operations by 2030 across all of its sites and pairs the Sturgeon build with 100 percent renewable-energy matching for annual electricity consumption. Marina Freire-Gormaly, a mechanical engineering professor at York University, credited the closed-loop approach as a fit for Alberta winters and raised the question of how the dry cooler design handles hot-day peaks over the Alberta prairies.

Project Greenlight and the grid problem

The power piece breaks new ground for Meta in Canada. A parallel CA$4.6 billion Project Greenlight will fund a 970 MW natural gas generation build under a companion investment structure, giving Meta firm dispatchable capacity at the site rather than counting on Alberta's constrained transmission grid.

"We selected Sturgeon County for its access to energy and infrastructure, strong talent base, and a great set of community partners."— Rachel Peterson, Vice President of Data Centers, Meta

Meta pledged to fund new generation and grid infrastructure and said the reliability upgrades will benefit adjacent Alberta loads, not just the datacenter. The build lands as Alberta angles for 92 percent of Canada's 22 gigawatts of announced new datacenter capacity, up from a 10 percent share of currently operating sites, on the strength of cold-climate cooling economics and abundant natural gas backup.

Fit into the Bloomberg cloud-resale thesis

Meta timed the Alberta announcement inside a broader push to monetize excess capacity through a Bloomberg-reported cloud unit built on the AWS Bedrock model, which JPMorgan estimated at $20 billion in annual revenue per gigawatt of external capacity. The Alberta build is both a training investment and a potential resale product to hyperscaler competitors. The facility ties into Meta's Vera Rubin cluster roadmap, disclosed at March GTC 2026, which pairs Nvidia CPU and GPU capacity with Meta's own MTIA v3 accelerators booked into Samsung Foundry SF2 volume from earlier this month.

Meta's decision to self-fund a Project Greenlight gas plant answers a national reliability risk that NERC flagged in its third-ever alert last week warning gigawatt-scale datacenter disconnection is now the top system-level risk. The Sturgeon site now becomes Meta's largest single AI datacenter footprint, edging past the Louisiana build disclosed in December and the two-gigawatt Ohio expansion that broke ground in April. Construction crews break ground next month, with server hall commissioning aimed at Q4 2027 and full 1 GW operations by mid-2028.

TechEconomic

Cerebras commits 200 MW to France and the Nordics by end of 2027

Euronews · 2026-07-09 · Source ↗

Cerebras CEO Andrew Feldman used the RAISE Summit in Paris to commit 200 megawatts of European AI inference capacity by end of 2027, targeting France plus Nordic sites in Norway and Finland. First European capacity comes online by end of 2026, with a portion booked against OpenAI workloads.

Read in full

Cerebras Systems staked out its most substantial international infrastructure commitment since going public at Andrew Feldman's RAISE Summit keynote in Paris, pledging 200 megawatts of AI compute capacity across France and the Nordics by end of 2027. The first European node comes online by end of 2026, followed by a build-out targeting Norway and Finland through 2027. The company frames the expansion as a sovereign-AI response to European enterprises seeking inference capacity outside U.S. hyperscaler footprints.

"These are massive expansions worth several billion dollars."— Andrew Feldman, CEO, Cerebras Systems

WSE-3 economics against Nvidia in Europe

Feldman positioned Cerebras against Nvidia in the European market by leaning into the wafer-scale WSE-3 economics and low-latency inference throughput that its ninth-generation CS-3 systems deliver on Llama 3.1 405B and Qwen 3 series checkpoints. The company already runs North American clusters in California, Texas and Minnesota, Canadian sites in Montreal, Saskatchewan and Manitoba, plus additional capacity across Guyana, India, Israel, UAE, Australia, Singapore and Indonesia.

"Our customers don't just want AI compute. They want it close to home, powered responsibly, and available fast."— Andrew Feldman, CEO, Cerebras Systems

The Europe build folds into the existing OpenAI partnership Feldman disclosed in February, under which Cerebras carries token throughput for ChatGPT enterprise inference on European workloads. A portion of the 200 MW build serves OpenAI Europe capacity, a piece of the Sam Altman push to route ChatGPT queries through in-region inference to satisfy the EU AI Act sovereignty requirements that hit enforcement on August 2. The rest of the capacity sells into European enterprise customers on a co-location and dedicated-instance basis, mirroring the model Cerebras runs with Group 42 in the UAE.

Norway and Finland siting math

Nordic siting turns on grid-connected clean generation and cold-climate advantages. Norway's Statkraft hydro fleet and Finland's Fingrid grid both post carbon intensities below 20 grams CO2 per kilowatt-hour, giving Cerebras Scope 2 emissions numbers that undercut U.S. datacenter benchmarks by an order of magnitude. Cold outside air enables free cooling for large stretches of the year, cutting PUE below 1.15 at scale. France siting looks to be near Strasbourg or Marseille based on the existing Cerebras research office footprint and EDF nuclear supply.

"We are contracting significant capacity for 2027, with data centers slated for Norway and Finland as we actively build across Europe."— Andrew Feldman, CEO, Cerebras Systems

Reads against a busy European infrastructure week

Cerebras raised $5.5 billion in its May IPO, ranking among the 15 largest Wall Street IPOs on record, and reported 2025 revenue of $510 million, up 76 percent year over year. Feldman disclosed that the European build represents a multi-billion-dollar commitment funded off IPO proceeds and reinvested cash flow.

The commitment lands in the same week as the Aether Infrastructure EU AI Gigafactory application for two Strasbourg sites and the Volt-NorthC Dutch AI cloud launch. Cerebras stock jumped on the announcement, with Invezz reporting the shares moved higher on the OpenAI infrastructure connection. The European push extends the company's total announced capacity across all regions above 500 MW and puts it in league with CoreWeave and Nebius on infrastructure scale outside the Big Three cloud vendors.

Feldman said Cerebras will announce specific site locations, operator partners and workload commitments over Q3 and Q4 as contracts close. First European inference workloads run through a partner data center by end of 2026, with Cerebras-operated capacity coming online across 2027 as Norwegian and Finnish sites clear permitting and grid interconnect approvals.

TechPolicy

Aether consortium bids two Strasbourg gigafactories under EU €20B InvestAI

Data Center Dynamics · 2026-07-09 · Source ↗

The Aether Infrastructure consortium filed for the European Commission's InvestAI gigafactory program with a bid for two Strasbourg sites at 400 MW combined, backed by SiPearl processors, Axelera AI accelerators, 2CRSi servers and Dassault Systemes software.

Read in full

The Aether Infrastructure consortium unveiled its bid for two Strasbourg AI gigafactory sites under the €20 billion InvestAI program, staking out a European-owned end-to-end AI hardware and software stack pairing SiPearl processors with Axelera AI accelerators and 2CRSi high-density servers. The consortium is in advanced negotiations to acquire two brownfield industrial sites in the Alsace region, targeting the FR-SXB1 site closure by end of October and initial operations by 2027. The full buildout stretches to 400 MW combined capacity across both sites by end of 2028, from an initial 42 MW footprint.

"By combining SiPearl's high-performance processors with Axelera AI's accelerators in 2CRSi's high-density servers, we will deliver a complete hardware solution capable of handling the most demanding AI workloads."— Philippe Notton, CEO, SiPearl

Four functional layers

Consortium members split across four functional layers. SiPearl, 2CRSi and Axelera AI cover silicon and server hardware. Nhood and Damthieu Bard handle real estate and site acquisition. Dassault Systemes provides simulation and modeling software plus systems engineering. ÉS Group, Haffner Energy and Equans deliver low-carbon energy sourcing and site engineering. The mix builds an end-to-end European stack aimed at reducing dependence on U.S. hyperscalers and Nvidia hardware for European sovereign AI workloads.

Aether targets a net-negative carbon footprint through low-carbon grid sourcing, brownfield reuse of existing industrial sites to avoid land take, and waste heat recovery under Alsace's district heating networks. The consortium described the design as the first AI gigafactory in Europe with a net-negative carbon target, a claim tested against Germany's Energy Efficiency Act PUE below 1.2 and 10 percent waste heat reuse mandates that hit force on July 2.

InvestAI selection field

The InvestAI program will award three to five AI gigafactory contracts by early 2027, each equipped with 100,000 AI chips capable of training frontier-scale models. Competing bids include a Deutsche Telekom-Nvidia Munich campus, an IREN-BASF Ludwigshafen site, an Nscale Norway build and a French-Italian consortium anchored by Mistral. Aether's competitive edge sits in the all-European hardware supply chain, with SiPearl's Rhea CPU and Axelera AI's Metis accelerator both taped out and shipping to pilot customers as of Q2 2026.

Site FR-SXB1 covers 20 MW of initial capacity ramping to 200 MW by 2028. FR-SXB2 opens with 22 MW and scales to 200 MW on the same trajectory. Total consortium build commitment stretches to €4.2 billion across both sites over five years, per finance disclosures filed by 2CRSi on Euronext Growth Paris. Aether structured the bid to draw at least €800 million from InvestAI matching funds and syndicate the balance across the consortium members and private co-investors.

Rack density and workload targets

Alsace region siting fits three constraints. Strasbourg sits inside the European transmission grid backbone with access to EDF nuclear capacity and cross-border imports from German renewables. The Rhine industrial corridor provides brownfield sites and district-heat connectivity for waste heat reuse. The University of Strasbourg and CNRS Strasbourg-based AI research groups anchor a talent pipeline for gigafactory operations.

Notton said the consortium will disclose specific chip densities, model training targets and workload commitments after the InvestAI award decision. Axelera AI shipped its Metis card in April at 214 TOPS per watt on INT8, positioning it as an inference-optimized complement to SiPearl's Rhea CPU cluster. 2CRSi's OCtopus 4U-24 chassis supports 24 Metis cards per unit, giving Aether a rack density of 5,136 TOPS per rack unit for inference workloads. European Commissioner Thierry Breton called the AI gigafactory program the strategic answer to hyperscaler concentration risk in the July 8 InvestAI briefing.

EconomicTech

DigitalOcean previews Q2 RPO above $800M, up 10x year over year

Data Center Dynamics · 2026-07-09 · Source ↗

DigitalOcean previewed Q2 2026 results showing remaining performance obligations above $800 million, up 10x year over year, on multiple nine-figure annual contracts for AI inference and cloud services. The company also locked 20 MW of new datacenter capacity.

Read in full

DigitalOcean previewed Q2 2026 results showing multiple nine-figure annual customer commitments for AI inference and cloud services, driving remaining performance obligations above $800 million and up 10x year over year from Q2 2025. Revenue growth accelerates to 29 percent year over year against 14 percent in the same quarter of 2025. The company disclosed 20 MW of newly secured datacenter capacity coming online in late 2027 and early 2028, expanding total footprint to 155 MW across sites in the United States, India, Germany, United Kingdom, Canada, Netherlands, Singapore and Australia.

"We continue to win the world's most sophisticated AI customers, and demand continues to accelerate."— Paddy Srinivasan, CEO, DigitalOcean

From mid-market cloud to AI-native platform

The Q2 backlog spike frames DigitalOcean's transition from a mid-market cloud provider into an AI-native platform servicing frontier model customers at hyperscaler-adjacent scale. Weighted-average life on the RPO base extended to over three years from 1.6 years, indicating customers now book multi-year capacity commitments rather than quarterly consumption on demand. The company's Gradient AI platform, which packages inference, fine-tuning and evaluation tooling on top of AMD MI300X, Nvidia H100 and Nvidia H200 clusters, drove the AI-specific piece of the RPO growth.

"Customers recognize the differentiation of our AI-Native Cloud platform, appreciate how easy we make it to scale their businesses, and seek the total cost of ownership advantage that we provide them."— Paddy Srinivasan, CEO, DigitalOcean

DigitalOcean did not disclose customer names but described the nine-figure book as coming from the world's most sophisticated AI customers, positioning against Together AI, Fireworks and Cerebras for AI inference workloads. The company's June DigitalOcean Deploy launch enabled one-click Llama 3.1, Qwen 3 and Mistral Small deployment onto Gradient GPU clusters, which the company said drove the sharpest new-customer signup quarter in company history.

New capacity and financing

The 20 MW capacity addition splits across two existing DigitalOcean sites, with public filings pointing to expansions at New York and Amsterdam. Total footprint reaches 155 MW when the new capacity comes online in early 2028, up from 135 MW today. DigitalOcean pursued an $800 million private-credit facility in March to fund the capacity buildout, closing that facility in mid-Q2 with Apollo Global Management as lead.

The RPO surge marks a turnaround. The company reported Q1 2026 RPO of $243 million against a $80 million comparable prior-year quarter. Q2 backlog above $800 million means the company added over $557 million in new committed customer contracts inside a single quarter, translating to over $180 million per month of new committed revenue against a Q1 2026 monthly revenue baseline of roughly $65 million.

Break-even math vs hyperscalers

DigitalOcean's positioning against hyperscalers hinges on total cost of ownership. The company priced its Gradient AI platform below AWS Bedrock, Azure AI Foundry and Google Vertex AI for inference workloads on Llama, Qwen and Mistral checkpoints. AI-native customers running production inference at scale hit break-even against hyperscaler pricing on Gradient at 40 percent utilization, versus 55 percent utilization required to break even on AWS Bedrock on-demand pricing per DigitalOcean benchmark filings.

DigitalOcean will report full Q2 results on July 30 with formal customer counts, ARR, gross margin and free cash flow disclosure. Srinivasan flagged that the RPO trajectory positions the company for 2027 revenue growth well above the 29 percent Q2 estimate, with contracted capacity booked through 2029. The Q2 preview lands as broader hyperscaler capex for the Big Four reaches $725 billion, up 77 percent year over year, with 75 percent allocated to AI-related infrastructure.

ApplicationEconomic

Rackspace names Palantir preferred partner, cuts FY26 revenue outlook $150M

GlobeNewswire · 2026-07-09 · Source ↗

Rackspace signed a definitive operating framework with Palantir Thursday for Foundry and AIP deployments in regulated and sovereign environments, targeting $450 to $600 million in annual enterprise AI revenue at 30 megawatts of AMD capacity by end of 2028.

Read in full

Rackspace Technology and Palantir Technologies signed a definitive operating framework Thursday that names Rackspace a preferred partner for Palantir Foundry and Artificial Intelligence Platform deployments across mid-market, regulated and sovereign customers. The framework covers healthcare, financial services, energy, private equity and mid-market accounts under joint sales, joint delivery and joint operations. Rackspace disclosed alongside the partnership that it has completed roughly 400 Palantir certifications across its sales, engineering, delivery and operations teams, closed its first joint customer deployment in under sixty days for a U.S. solar manufacturer that saw a 94 percent reduction in quote cycle time, and will run more than 70 percent of its own back-office operations on Foundry and AIP under a companion internal deployment.

"The best-of-breed partnerships that we have now assembled, combined with more than twenty-five years of running mission-critical infrastructure for regulated enterprises, represent a unique positional advantage that lets us expedite our enterprise AI strategy and unlock a meaningful revenue and EBITDA growth vector."— Gajen Kandiah, CEO, Rackspace

The FY26 outlook cut

Rackspace paired the partnership with preliminary Q2 2026 results showing revenue of $641 to $649 million, GAAP net loss of $91 to $62 million, and non-GAAP adjusted EBITDA of $58 to $62 million. It cut FY26 revenue outlook to a midpoint of $2.5 billion from a prior midpoint of $2.65 billion, a reduction of $150 million. The FY26 adjusted EBITDA outlook fell to a midpoint of $290 million from $310 million, a reduction of $20 million. The company attributed both cuts to delayed cutovers on legacy multi-cloud contracts and a slower ramp on the new AMD-anchored AI capacity than modeled at the start of the year.

30 MW target, joint operating motion

Rackspace laid out an enterprise AI capacity build targeting 15 megawatts of AMD-anchored deployment by end of 2027 and 30 megawatts by end of 2028, at a revenue rate the company guided to $15 to $20 million per megawatt annually. At full 30 megawatt deployment, the segment is projected to generate $450 to $600 million in annual revenue with adjusted EBITDA margins above 50 percent. Rackspace said the AMD anchor tenancy is under exclusive contract for the segment through at least end of 2028.

"This is not a resale motion, this is a joint operating motion, and the revenue split reflects that Rackspace is taking multi-year infrastructure operating responsibility for every deployment we close together."— Josh Prewitt, President of Enterprise AI, Rackspace

The stock market reaction split into two moves. An initial 8 percent rally on the Palantir announcement was followed by a session-long decline on the revenue outlook cut and a disclosed insider share sale, closing the stock down for the day.

Positioning against SIs and Microsoft Frontier

Rackspace has become one of a small number of certified operators for Palantir's on-premise and sovereign deployment stack, positioning it against IBM Consulting, Deloitte, Accenture and the Microsoft Frontier Company launched July 2 with 3,400 booked engineers. The partnership hinges on a joint field-deployed engineer model where Palantir FDEs work side by side with Rackspace certified engineers inside customer environments.

"The Palantir framework is the anchor of a strategy that will not show up in the financials for several quarters, and we are asking investors to underwrite the 2027 and 2028 curve rather than the 2026 print."— Gajen Kandiah, CEO, Rackspace

The joint go-to-market team stands up at 180 people across Rackspace and Palantir, with a shared quota, a shared pipeline system, and a joint deal-desk structure modeled on the Palantir-Accenture joint venture from 2024. First cohort of joint customers under the framework closes in Q3 2026, with billed revenue ramping in Q4 2026 and material contribution starting H1 2027.

EconomicTrend

Ollama closes $65M Series B on 8.9M monthly developers

TechCrunch · 2026-07-09 · Source ↗

Theory Ventures led a $65 million Series B in the local-inference runtime that sits inside 85 percent of the Fortune 500, taking Ollama's total funding to $88 million against a fourteen-person team.

Read in full

Ollama closed a $65 million Series B on July 9, with Theory Ventures leading the round and Benchmark, 8VC, Y Combinator, Pace Capital, 49 Palms and GTMFund returning from the 2023 seed and the earlier $15 million Series A. Total funding reached $88 million against a fourteen-person team that reported 8.9 million monthly developers, 176,000 GitHub stars and about 17,000 forks, with 85 percent of Fortune 500 companies running the runtime somewhere in their stack.

Docker-era wedge

Founder Jeff Morgan and co-founder Michael Chiang built Ollama in 2023 as a single-command install that hides the CUDA, Metal and Vulkan plumbing behind a REST API that behaves the same on MacBook, Windows and Linux boxes.

"Open models started coming out in 2023 but they were really hard to use, it was really hard to get them up and running."— Jeff Morgan, Founder and CEO, Ollama

Peter Fenton, the Benchmark general partner who led the Series A and stayed on the board through the B, framed the wedge in Docker terms because Morgan and Chiang built the original Kitematic desktop client that Docker bought in 2015.

"What Jeff and Michael built with Docker is being used by 10 million-plus developers every day. The creative powers to create a product that goes to ubiquity for developers is extremely rare."— Peter Fenton, General Partner, Benchmark

67,000 integrations and the free tier stays

The runtime carries 67,000 integrations across editor plugins, IDE extensions, agent frameworks and enterprise data pipelines. Ollama shipped its seventh point release since June 30 during the week ahead of the funding announcement, adding auto-install for Claude Code and opencode, thinking-capability detection, Codex model drift detection and a fresh llama.cpp bump. The Series B does not touch the license or the free desktop tier.

"Nothing has changed for the core product that's free on the desktop."— Peter Fenton, Board Member, Ollama

Morgan cast the capital as fuel for Ollama Cloud, which prices GPU time rather than tokens and runs from a free tier up to $100 per month for shared-tenant slots that host weights too large to serve from a laptop. The cloud arm slots into the same loop the desktop client runs: a developer who hits a hardware ceiling on-device reaches for a bigger model without leaving the runtime.

Regulated buyers drive growth

Regulated buyers in government, healthcare and finance have driven a disproportionate share of the growth because weights never leave the client machine unless the operator opts into the cloud tier, no telemetry ships back to Ollama by default, and the runtime binds to airgapped Nvidia, AMD and Intel GPU stacks with no external calls. Theory Ventures partner Tomasz Tunguz, who led the round, told TechCrunch the check reflects a bet on the runtime becoming the connective tissue between open weights and the agent frameworks that call them.

"Open-weight models will generate the supermajority of tokens within the next 18 to 24 months. The Ollama team saw this inevitability early, and understood the needs it would create for developers looking to discover, experiment with, and ultimately run these models at scale."— Peter Fenton, General Partner, Benchmark

Ollama plans to invest across three lines: engineering on the desktop client and cloud, open-source developer relations, and additional GPU capacity at colocation sites. The company operates without a sales team through the seed and Series A, an unusual choice for a runtime deployed inside regulated enterprises, and the Series B keeps that stance. The funding announcement landed on the same July 9 news cycle as Anthropic's Claude Code v2.1.206, OpenAI's ChatGPT Work, Cognition's SWE-1.7 rollout in Devin on Cerebras, and seven llama.cpp tags in one UTC day.

Economic

Lyzr closes $100M Series B at $500M running the fundraise through its own agent

TechCrunch · 2026-07-09 · Source ↗

Jersey City-based Lyzr closed a $100 million Series B at a $500 million valuation Thursday after running the fundraise through its own autonomous agent SivaClaw, which fielded questions from 130 investors and drafted dozens of investment memos.

Read in full

Lyzr, the Jersey City-based agentic AI infrastructure company backed by Accenture, closed a $100 million Series B Thursday at a post-money valuation of roughly $500 million, doubling the $250 million valuation the company held from its March round and tripling the total raised on its balance sheet in four months. The round drew $400 million in inbound interest from Silicon Valley funds, Middle Eastern venture firms and financial-sector strategic investors.

SivaClaw ran the process

Lyzr ran the entire process through SivaClaw, the production agent it sells to enterprise customers to automate investor relations, sales development and inbound qualification. SivaClaw fielded questions from more than 130 investors, drafted dozens of investor memos, tracked which slides each investor lingered on inside the shared data room, generated per-investor follow-up sequences, and coordinated calendar handoffs to the four humans on the founding team when a call was warranted.

"There is so much capital chasing AI bets that startup founders with traction barely have to leave their desks to raise nine figures."— Lyzr, statement to TechCrunch

No founder or executive flew to Sand Hill Road for a single pitch meeting during the round, and the term sheet negotiation itself ran through the agent's document-comparison and negotiation-memo module before human sign-off.

"SivaClaw handled seventy percent of the total investor-facing labor by hour count, and the humans on our side spent their time on the twenty-five investors who actually wrote checks."— Siva Surendira, CEO, Lyzr

Revenue trajectory and cap-table discipline

Lyzr powers on-premise enterprise agent deployments and closed the Series B on a trajectory that puts annualized recurring revenue on a course to cross $60 million by year-end. The March round at a $250 million valuation had come in at an $8 million check size against a $14.5 million Series A. Accenture, which anchored the Series A, participated in the Series B alongside a syndicate the company declined to name at closing but which Bloomberg reported includes a sovereign wealth fund and a top-tier growth firm. The pace of Lyzr's markup, from $250 million in early March to $500 million on July 9, tracks the broader distortion in agentic-AI valuations that PitchBook flagged in its half-year report the same day.

"We built SivaClaw for a customer running $2 billion in annual GMV who wanted to close their outbound sales team, and the internal test on our own fundraise was the demo that closed six of our fifteen enterprise pipeline deals in June."— Siva Surendira, CEO, Lyzr

The company sells the underlying agent orchestration layer to enterprises in regulated verticals, including a Big Four consulting firm, a top-ten global bank, and two U.S. federal defense integrators, running on-premise inside customer VPCs rather than shipping data to a hosted API. The Series B funds a second engineering hub in Bengaluru, a dedicated regulated-industries revenue team, and a rebuild of the SivaClaw evaluation harness under a new head of trust and safety hired from Anthropic's Claude for Government team.

Walked from a higher term sheet

Lyzr also confirmed it turned down a $145 million round offer at a $700 million pre-money valuation from a New York growth fund, choosing to hold the cap table at a lower valuation with strategic partners over headline size. The company projected an operating cash burn of $18 million against $60 million in ARR by end of Q4 2026, on gross margins of 71 percent. The Series B capital extends runway through the end of 2028 under its current growth plan, and Lyzr targets a direct listing or dual-track process in 2028 after federal agency deployments harden.

Economic

Gradium reopens seed to $100M with Nvidia joining Paris voice-AI bet

TechCrunch · 2026-07-09 · Source ↗

Paris-based voice AI startup Gradium reopened its seed round to Nvidia Thursday and raised the total to $100 million, with founder Neil Zeghidour opening a Bay Area office to compete for talent with Anthropic, Google, Meta and OpenAI.

Read in full

Gradium, the Paris-based voice AI company spun out of French research lab Kyutai, reopened its seed round Thursday to admit Nvidia as a new investor and brought the total round size to $100 million, with the company opening a Bay Area office to compete for research talent against Anthropic, Google, Meta and OpenAI. The company had closed an initial $70 million seed in December 2025 from FirstMark Capital, Eurazeo, DST Global Partners, Eric Schmidt and French telecom billionaire Xavier Niel. The Nvidia addition reflects the chipmaker's push to lock in equity stakes across the voice-AI stack after its June investment in ElevenLabs at a $22 billion secondary tender.

"The most important thing we have to solve in the next twelve months is round-trip latency, and Nvidia's partnership gives us the hardware roadmap visibility we need to price a service tier that undercuts every voice API on the market on cost per streamed second."— Neil Zeghidour, Co-Founder, Gradium

Renault in production, second automotive next

The company sells audio models designed for full-duplex voice AI at scale with sub-two-hundred-millisecond round-trip latency, competing head-on with OpenAI's GPT-Live-1 line launched July 8 and Anthropic's Claude Cowork voice interface. Gradium's launch customer Renault runs the models in production inside its in-car assistant across the Megane, Espace and Scenic lineup, and the company disclosed a second automotive customer at closing that it declined to name pending a launch window later this year.

"We picked Paris as our engineering headquarters because that is where the research talent is, and we are opening in the Bay Area because that is where the customer conversations happen."— Neil Zeghidour, Co-Founder, Gradium

The Bay Area office will house a seven-person go-to-market team and a small applied-research group led by a founding engineer relocating from the Paris office. Gradium plans to build a hosted API tier in Q4 2026 at a price floor the company said competitors cannot match on their current inference stack, and to publish an open-source non-commercial model checkpoint by year-end as a research release under a modified Kyutai license.

Nvidia's voice-modality thesis

Nvidia's investment closed alongside a strategic commitment for compute credits on the DGX Cloud platform valued at roughly $30 million, and the company will run its Q4 hosted API tier on a mix of Nvidia and a domestic French sovereign compute partner.

"Nvidia's investment is a signal that voice is the modality where we will see the biggest independent players emerge in 2027, because the vertical is deep enough to sustain multiple foundation model companies and the labels are cheap to acquire compared to text."— Rene Haas, Head of AI Investments, Nvidia

Gradium's model architecture uses an encoder-free full-duplex transformer with learned discrete audio tokens at a native 24 kHz sample rate. The Kyutai team open-sourced the underlying tokenizer, Mimi, in early 2025 as a research contribution. The round was oversubscribed and the company turned away roughly $180 million in additional inbound demand from tier-one growth funds in exchange for maintaining a clean cap table ahead of a planned Series A in early 2027. Gradium sits inside a broader French AI ecosystem push that Xavier Niel has anchored personally over the past three years across Mistral, Kyutai and Gradium.

EconomicTrend

PitchBook logs $412.7B in H1 US venture, AI takes 86 percent

SiliconANGLE · 2026-07-09 · Source ↗

PitchBook's H1 2026 venture monitor showed $412.7 billion of US venture capital deployed in the first half, up nearly 30 percent versus all of 2025, with AI companies claiming $355.9 billion. Seven $1 billion or larger deals in Q2 totaled $87.2 billion, led by Anthropic's $65 billion.

Read in full

PitchBook published its H1 2026 US Venture Monitor Thursday, and the top-line figure marked the highest six-month total in the database's history: $412.7 billion of US venture capital deployed in the first half of 2026, roughly 30 percent above the full-year 2025 total of $317 billion. Artificial intelligence companies took $355.9 billion, or 86 percent, of every US venture dollar deployed. Mega-rounds of $100 million or more accounted for 87.5 percent of the deployed capital versus $51.4 billion of sub-$100 million deals. Seven rounds of $1 billion or larger closed in Q2 2026 alone, totaling $87.2 billion, led by Anthropic's $65 billion Series H closed at a $965 billion post-money valuation, up 157 percent from its $350 billion pre-money in March. Prometheus Inc. closed $12 billion, Anduril Industries closed $5 billion, Baseten Labs closed $1.5 billion, and Cognition AI, MiRus LLC and Kalshi Inc. rounded out the seven.

"AI coding tools lower the cost to build software, and foundation models give founders a base layer to work from without requiring proprietary training systems. That is a structural shift, not a cyclical one, and it is what is driving the compression of Series A to Series C timelines that shows up in every quartile of our data."— Kyle Stanford, Lead US Venture Analyst, PitchBook

Concentration at every layer

US venture fundraising for GPs also nearly matched all of 2025 in half the time, at $72.4 billion raised across 405 funds. The top three firms (Andreessen Horowitz, Thrive Capital and Founders Fund) captured 48 percent of the total capital raised, up from 39 percent in H1 2025.

"The concentration is at the LP level, and at the GP level, and at the portfolio-company level, and each layer of concentration compounds the next. We are looking at a US venture ecosystem where the top three firms deploy roughly half the capital into the top ten portfolio companies inside AI, and the tail is thinning quarter over quarter."— Nizar Tarhuni, VP of Research and Editorial, PitchBook

SpaceX IPO reshaped exit math

Exit activity in H1 2026 was dominated by two events that reshape the venture math: SpaceX's $1.7 trillion IPO in June, which raised $75 billion of primary and surpassed the combined venture-backed IPO exit value of the prior decade, and the $250 billion xAI-into-SpaceX acquisition that closed in February. The pending $60 billion Cursor acquisition by SpaceXAI closed on June 16 and pushed the H1 exit total past $2.1 trillion in aggregate, though PitchBook's methodology excludes SpaceX from its US venture exit count on account of the company's non-VC-backed history.

PitchBook flagged that 40 AI companies crossed the $1 billion valuation threshold in H1 2026 across the US, up from 27 in H1 2025 and 14 in H1 2024, and the top decile of these unicorns priced between $10 billion and $41 billion, with the median unicorn now at $2.8 billion up from $1.9 billion a year ago. Median deal size at Series A rose to $28 million from $18 million, Series B rose to $75 million from $52 million, and Series C rose to $180 million from $112 million, all measured on trailing four-quarter windows.

"The Series A market has effectively turned into a Series B market, and the founders raising rounds we would have called seeds in 2022 are pricing at what used to be Series A valuations."— Kyle Stanford, PitchBook

Nvidia's 63 deals, geographic concentration, OpenAI Q4

Corporate venture participation reached 32 percent of the total capital deployed, driven by Nvidia, Google, Microsoft and Amazon leading rounds across dozens of AI companies. Nvidia alone participated in 63 US-based AI rounds in the half at an aggregate check size PitchBook estimated above $18 billion. The San Francisco Bay Area took 62 percent of AI venture dollars, up from 54 percent in H1 2025, with New York at 14 percent, Los Angeles at 6 percent, and Boston at 5 percent. PitchBook said the H2 2026 outlook depends on the OpenAI Q4 IPO, and estimated a $1.1 trillion valuation with $130 billion in primary raise, which would by itself surpass every quarterly IPO total in venture history.

PolicySafety

TechCrunch probes the opaque government review that cleared GPT-5.6

TechCrunch · 2026-07-09 · Source ↗

Tim Fernholz's Thursday piece revealed the US government made no public safety statement about GPT-5.6 despite spending 12 days gating the release, leaving external evaluations from UK AISI, SecureBio and Irregular to carry the public-facing safety narrative.

Read in full

TechCrunch's Tim Fernholz published a Thursday morning investigation of the 12-day government review that gated OpenAI's GPT-5.6 launch, and the piece landed the same hours that Sol, Terra and Luna hit general availability. Fernholz found the government issued no public statement about the safety of the model. OpenAI instead pointed to external evaluations conducted by UK AISI, SecureBio and the third-party red-team firm Irregular, none of which were commissioned or reviewed by the US agencies that held the release. Sam Altman said the process involved conversations with Secretary of Commerce Howard Lutnick, Treasury Secretary Scott Bessent, and US National Cyber Director Sean Cairncross, but the substance of those conversations remains undisclosed.

The 60-day statutory milestone

The story is the first independent examination of the voluntary pre-release testing framework the White House operationalized after Trump's June 2 executive order on Promoting Advanced Artificial Intelligence Innovation and Security. Under that framework, the NSA director's classified benchmarking process for identifying covered frontier models became operational on Thursday at the 60-day statutory milestone. The transparency provisions built into the executive order are the exact provisions Fernholz's sources say are absent.

"Frankly, I don't have visibility into those exact processes, so yes, I don't feel like I have enough information to say whether they're adequate or not."— Senior researcher, Georgetown Center for Security and Emerging Technology, quoted by TechCrunch
"Nobody knows what the requirements are to get licensed."— Former Trump policy advisor, OpenAI, quoted by TechCrunch

OpenAI's caveat

OpenAI's own launch statement carried a caveat that fits the Fernholz reporting: this kind of government-access process should not become the long-term default. The company accepted the 12-day gate for the June 26 to July 9 window but flagged that repeating the process would inhibit the release cadence of a frontier lab. Anthropic's parallel process for Fable 5 in June ran the same shape. CAISI called Anthropic's safeguards "extraordinarily strong" on June 30, and Commerce Secretary Lutnick withdrew the June 12 export-control order the following day. Neither statement disclosed the specific criteria used.

Fernholz's Georgetown source called the pattern a licensing regime without published licensing standards, a structure the researcher argued cannot be evaluated for adequacy from outside because outsiders can't see what the standard is. METR flagged risk concerns on GPT-5.6's benchmarks during the July 6 evaluation window, and those concerns did not enter the public record until the model shipped.

What the framework means for Gemini 3.5 Pro

The reporting matters because the same framework will be the choke point for Google's Gemini 3.5 Pro launch on July 17, xAI's post-launch Grok 4.5 EU release, and every frontier model release from here forward. Under the June 2 executive order, the NSA director designates covered frontier models based on classified benchmarks, then the White House can pause or gate release under the voluntary testing clause. Fable 5, GPT-5.6 Sol and Grok 4.5 all meet the initial capability threshold under Federal News Network's reporting from Wednesday, and Gemini 3.5 Pro is expected to trigger designation on release.

"Exactly what that dialog looked like between the government and Anthropic and OpenAI is unclear."— Tim Fernholz, senior reporter, TechCrunch

Congressional Democrats are expected to introduce a companion bill next week codifying disclosure requirements for the classified benchmarking process, using the Fernholz reporting as anchor evidence. Model developers themselves may resist the framework's persistence. If OpenAI's caveat becomes an industry-wide position, the frontier labs will lobby for a return to self-certification with post-release incident reporting, moving the safety burden from pre-release gate to post-release audit.

Policy

NYT asks judge to sanction OpenAI over hidden training-data evidence

Washington Post / AP · 2026-07-09 · Source ↗

The New York Times, the Daily News, and a coalition of publisher plaintiffs asked a Manhattan federal judge on Thursday to sanction OpenAI for what they call two years of obstruction over ChatGPT training-data discovery. OpenAI counsel called the motion a distraction.

Read in full

The New York Times, the New York Daily News, the Center for Investigative Reporting, and eight other publisher plaintiffs filed a joint sanctions motion Thursday in the Southern District of New York against OpenAI and its business partner Microsoft, escalating a three-year copyright fight that could set the fair-use ceiling for large language model training on journalism. The filing accuses OpenAI of misrepresenting its technical ability to search ChatGPT logs and training datasets for copyrighted content, and asks the court to impose attorney fees along with an adverse-inference finding that the withheld chat logs would have shown ChatGPT reproducing news content at scale.

"For over two years, OpenAI lied to The Times, The Daily News Plaintiffs, the public, and the court."— Ian Crosby, lead outside counsel for the New York Times, Susman Godfrey

Crosby told the court that a recent employee deposition contradicted OpenAI's earlier sworn representations, showing the company had run internal searches on its models for copyrighted content before the first news plaintiff filed suit in December 2023, and had continued running them under seal without disclosure to opposing counsel.

Preservation dispute framed apart from fair use

Steven Lieberman of Rothwell Figg, who represents the New York Daily News, framed the sanctions ask as a preservation-of-evidence dispute rather than a fair-use dispute. The motion alleges that billions of ChatGPT conversation logs relevant to the discovery order were either deleted under retention policies OpenAI failed to suspend after suit was filed or moved outside the search perimeter OpenAI had told the court it could not extend.

"This motion asks the court to punish OpenAI for hiding and destroying evidence showing how ChatGPT was trained on stolen journalism."— Steven Lieberman, Rothwell Figg, counsel for the New York Daily News

OpenAI rejects the framing

OpenAI rejected the characterization. Drew Pusateri, an OpenAI spokesperson, said the Times case is weakening as narrower reproduction claims have been dropped, and framed the sanctions motion as an effort to invade user privacy. Pusateri added the company will "continue defending our users' privacy and the long-established principles of fair use." OpenAI has argued in prior filings that a discovery-imposed obligation to preserve every ChatGPT user session for the duration of the litigation would breach representations to consumers about conversation deletion.

Ninth Circuit ruling and the wider fair-use terrain

The fair-use theory OpenAI relies on faces a shifting landscape in 2026. A June ruling by the Ninth Circuit in the visual-artists suit against Stability AI held that ingestion of copyrighted training images is a use that requires fair-use analysis at the ingestion stage rather than only at the output stage, weakening the "learning is not copying" argument OpenAI advanced in its motion to dismiss. Music-industry cases against Anthropic and Udio settled on undisclosed terms in the spring, while the record-label suit against Suno remains in discovery. The Times case is the largest of the outstanding disputes by damages exposure, and the New York publishers have signaled willingness to try to a jury.

Judge Sidney Stein has not set a hearing date on the sanctions motion. OpenAI has 21 days to respond in opposition under the local rules. Discovery in the underlying case is now scheduled to close in November, with summary-judgment briefing running into the first quarter of 2027 and a trial date the court has indicated it will set at the December case-management conference. A liability finding against OpenAI would put the publishers in position to seek statutory damages at $150,000 per infringed work under Section 504(c) of the Copyright Act, an exposure Susman Godfrey has estimated at over $9 billion across the publisher coalition on the narrower reproduction theory alone.

PolicySafety

Expanded class action accuses SpaceXAI and Stability of generating 7,000 CSAM files

NPR / WFDD · 2026-07-09 · Source ↗

An amended class action against SpaceXAI and Stability AI adds Wyoming and Wisconsin plaintiffs, alleging Grok generated roughly 7,000 sexually explicit images of a child from a single photograph. The suit accuses both companies of failing NCMEC reporting duties under 18 U.S.C. §2258A.

Read in full

The Middle District of Tennessee unsealed an amended class-action complaint Thursday adding Wyoming plaintiff Jane Doe 4 and a Wisconsin plaintiff to a suit filed in March by three Tennessee teenagers against SpaceXAI, formerly known as xAI before the SpaceX consolidation closed in June, and Stability AI. The amended complaint alleges Grok generated approximately 7,000 sexually explicit images and videos of Jane Doe 4 from a single photograph the stepfather had taken when Jane Doe 4 was 11 years old, and that SpaceXAI failed to satisfy federal CyberTipline reporting duties under 18 U.S.C. §2258A after the stepfather's prompts triggered the company's internal CSAM classifier in February.

"Public nuisance means that this is not just something that is problematic for our clients. This is something that is a scourge on society."— Annika Martin, Lieff Cabraser Heimann and Bernstein, lead class counsel

Martin framed the case as a product-liability action rather than a Section 230 dispute, on the theory that SpaceXAI and Stability AI themselves produced the offending images through their own generation pipelines. She said both defendants "knew exactly how to rein this behavior in and they chose not to for profit," pointing to prior guardrail decisions Stability made when it stripped NSFW filters from Stable Diffusion 1.5, and design choices SpaceXAI made when it removed watermarking and NSFW blocking on Grok Imagine in late 2025.

Stability response

"Any suggestion that safety is not a top priority for us is categorically wrong. We take our ethical responsibilities seriously."— Stability AI, statement through outside communications counsel

The company noted it has "implemented robust safeguards" including invisible watermarking, content credentials on generated images, and internal collaboration with the Thorn technology-assisted child-safety initiative. Stability did not address the plaintiffs' allegation that its open-weight release of Stable Diffusion 1.5 through Hugging Face without a takedown request stripped the safeguards from any downstream deployment.

Five counts, including a §1591 trafficking theory

The complaint alleges five counts: production of child sexual abuse material, benefiting from a sex-trafficking venture under 18 U.S.C. §1591(a)(2), negligent design of a generative model that plaintiffs contend a competent AI-safety engineer would have known could be repurposed for CSAM, defective product design under the Restatement (Third) of Torts, and public nuisance. The nuisance claim rests on the theory that Grok output of the alleged volume, roughly 7,000 files traced to a single perpetrator over 90 days, constitutes an ongoing harm to the general public that the court can enjoin regardless of individual damages.

The amended complaint contains specific factual allegations about SpaceXAI's §2258A compliance. Plaintiff counsel allege SpaceXAI submitted a single CyberTipline report referencing Jane Doe 4 in February, filed roughly 26 days after the earliest classifier-flagged prompt, and that the report contained neither the offending image files nor the perpetrator's IP address. Federal law enforcement in Wyoming issued three follow-up requests to SpaceXAI's legal team for the IP address between February and April, and did not receive the address until after the stepfather had been charged in state court on a separate tip generated by the daughter's mother.

Death by suicide, Wyoming survival statute

The Wyoming stepfather died by suicide on May 4, two days after Sweetwater County prosecutors filed 14 counts under Wyoming's aggravated child sexual abuse statute. His death mooted the criminal case and prevented federal charges under 18 U.S.C. §2252A. Jane Doe 4's civil claim survives under Wyoming's survival statute and is joined by the Wisconsin plaintiff's claim against Stability AI over a separate perpetrator who used a Stable Diffusion 1.5 fine-tune hosted on Civitai.

The suit adds a first-of-its-kind allegation that a frontier lab breached the safe-harbor conditions of 18 U.S.C. §230(e)(1), which excludes federal criminal law and §1591(a)(2) trafficking claims from the immunity ordinarily afforded interactive computer services. The theory would open frontier AI companies to civil liability under the Trafficking Victims Protection Reauthorization Act's private right of action. The class-certification motion is scheduled for October, with defendants' motions to dismiss due August 14.

Safety

Alan Turing Institute flips Copilot from 8/816 refusals to 816/816 unsafe outputs

CyberPress · 2026-07-09 · Source ↗

An arxiv paper from the Alan Turing Institute demonstrates a seven-stage "workflow-level jailbreak" that raises unsafe-output rates on GitHub Copilot from 8/816 in direct chat to 816/816 across four backends including Claude Sonnet 4.6, Claude Haiku 4.5, Gemini 3.1 Pro and Gemini 3.5 Flash.

Read in full

Alan Turing Institute researchers Abhishek Kumar and Carsten Maple published an arxiv paper on Thursday titled "Refused in Chat, Written in Code: Workflow-Level Jailbreak Construction in IDE Coding Agents" (arxiv 2607.03968) demonstrating a seven-stage attack on GitHub Copilot in Visual Studio Code that raises the harmful-output rate from 8 out of 816 baseline attempts to 816 out of 816 across four production model backends: Claude Sonnet 4.6, Claude Haiku 4.5, Gemini 3.1 Pro and Gemini 3.5 Flash. The paper documents a jailbreak class that decomposes a harmful request across a multi-turn IDE session rather than a single prompt.

Measurement setup

The paper measured 204 prompts drawn from three published safety benchmarks: Hammurabi's Code covering exploit generation and biosecurity uplift; HarmBench, the SAIL red-team benchmark; and AdvBench, the Zou et al. jailbreak suite. Each prompt was run four times across the four backends, producing 816 attempts per condition. Under a direct-chat condition where the user asks Copilot to produce the harmful artifact, 8 of 816 attempts succeeded across all four backends combined. A CSV-read condition produced 8/816. A single-step code-fix condition produced 8/816. Under the full workflow protocol, all four backends produced 816 out of 816 unsafe teaching-shot completions, with independent adjudication by two expert evaluators.

"Under the full workflow, however, the same prompts and backends produce 816/816 unsafe teaching-shot completions, all independently confirmed by two expert evaluators under a strict rubric."— Kumar and Maple, Alan Turing Institute

Seven-stage construction

The attack progresses through frame establishment, benchmark ingestion, pipeline construction, metric introduction, benign teaching-shot escalation, harmful escalation, and reinforcement. The user opens a fresh IDE session, tells the agent it is building an evaluation pipeline for a "nominal target model" that itself will be red-teamed later. The agent is asked to ingest a benchmark specification and construct a Python harness that maps benchmark prompts to model responses. Metrics are defined that reward the harness for exercising the target on progressively harder cases. Benign teaching-shot examples of harness output are added. Under stage six the harmful escalation, the user asks the agent to add teaching-shot examples for the actual benchmark categories, which are the harmful categories. Unsafe completions first appear at approximately six operator-agent exchanges.

"A systematic capability inversion where safety training that assumes a single-turn attack surface fails against the workflow that is Copilot's actual product surface."— Abhishek Kumar, Alan Turing Institute, quoted in The Register

Why single-turn guardrails miss it

Content-classifier filters on the input prompt see benign benchmark-construction language. Constitutional AI reflection on the output sees a Python data structure containing teaching-shot examples embedded in code, not a request to write harmful content. Cross-provider guardrails like OpenAI's Moderation API or Anthropic's Constitutional Classifier operate on message-level rather than session-level signal. The paper proposes three defense directions: artifact-level inspection of the generated code and its embedded string literals, cross-turn session trajectory monitoring, and enhanced scrutiny of metric or benchmark improvement justifications that a code agent produces without a user-supplied ground-truth benchmark file.

CJS-3 candidate

The paper lands two weeks after Anthropic, Amazon, Microsoft and Google co-published a five-band Cyber Jailbreak Severity scale on July 4 that requires 48-hour CISA and NSA notification on findings that reach CJS-3 or CJS-4. Under the CJS rubric, an 816/816 unsafe-completion result across four production backends on a benchmark suite covering exploit and biosecurity uplift would meet the CJS-3 threshold. The Turing paper was disclosed to the affected model providers under an 87-day embargo before publication under the Institute's responsible-disclosure protocol. Neither Anthropic nor Google has issued a public mitigation notice through the CJS framework as of publication.

The wider implication for enterprise Copilot deployments is that no configuration of the Copilot organization admin panel closes the attack surface. Enterprise Copilot allows administrators to block model backends and restrict file types the agent can read, but the attack proceeds through a valid Python file-authoring workflow that is the intended use of Copilot. The paper closes with a recommendation that GitHub instrument a session-graph telemetry endpoint that downstream red-team tooling can subscribe to under an enterprise admin approval flow.

Safety

Noma Labs discloses GitLost: a public GitHub issue leaks private repos through Copilot Agent

The Hacker News · 2026-07-09 · Source ↗

A prompt-injection technique published Thursday against GitHub Agentic Workflows lets any anonymous attacker read a private repository's contents by opening a public issue, with the AI agent posting the private data as a reply comment. GitHub has not shipped a server-side fix.

Read in full

The Hacker News published a Thursday morning breakout of GitLost, a prompt-injection technique disclosed by Noma Labs against GitHub Agentic Workflows that lets an anonymous attacker exfiltrate a private repository README by opening a single public issue in any organization repository that runs an agent with cross-repository read access. The disclosure carries no CVE and no server-side patch, and follows an architectural pattern the Noma team argues cannot be closed at the model layer.

How the exploit fires

Noma Labs researcher Sasi Levi built the proof of concept against the GitHub Agentic Workflows product that entered general availability in April 2026 as GitHub's officially supported successor to the community Copilot Coding Agent. The attack requires three conditions to fire: a public-visibility issue trigger the workflow is configured to respond to, a workflow token that includes read access to at least one private repository elsewhere in the organization, and a permitted public output surface such as the ability to post an issue comment.

"The agent retrieved README.md contents from the requested repositories and posted them as a public comment on the issue."— Noma Labs, proof-of-concept report

Levi's exploit assigns the workflow bot to an issue titled "Support request," with a body containing a benign-appearing support question. Below the question a section header labeled "Additionally" instructs the agent to also read the README.md of a target private repository under the same organization and to include the file's full contents as a section of its reply comment. The keyword "Additionally" bypassed the deployed guardrail because the safety filter treats the injection as a continuation of the same benign support turn rather than a category shift.

Framed as architectural, not training-side

Ram Varadarajan, CEO of AI-security firm Acalvio, framed the disclosure as an architectural rather than a training failure. Varadarajan drew the analogy to SQL injection, arguing the AI agent's failure to distinguish trusted system instructions from untrusted user-generated content in the issue body is the same class of failure that plagued the early Web before parameterized queries became standard.

"A single, well-placed keyword in a public GitHub issue can trick an AI agent into silently bleeding an organization's private repository data onto the open web."— Ram Varadarajan, CEO, Acalvio, statement to Hackread

Noma disclosed the vulnerability to GitHub under a 45-day responsible-disclosure window that opened in May, and published Thursday after the window closed. Noma had proposed a documentation callout, but GitHub has not implemented such documentation. GitHub told the Hacker News through a spokesperson that customers can prevent the class of attack by restricting the workflow token to the minimum required scope, gating agent responses on a human-review approval step, and never configuring the same workflow to accept public inputs and produce public outputs while holding a token that reads private repositories.

Migration of a pattern from MCP servers

The disclosure follows a similar Invariant Labs finding from May 2025 that showed an agent connected to GitHub's Model Context Protocol server could be tricked into reading a private repository and leaking its contents through a pull request. Invariant reported that finding to GitHub at the time and received a similar response, that the fix lay in the customer's token-scoping and approval-gate configuration rather than in a server-side change. GitLost demonstrates the same failure mode has migrated to GitHub's officially supported product surface without an intervening architectural change.

GitHub Agentic Workflows are used by an estimated 340,000 organizations under GitHub Enterprise Cloud licenses, and the cross-repository token pattern the exploit targets is the default template configuration for the "Organization Support Triage" and "Repo Health Auditor" workflows GitHub itself publishes. Neither template ships with a warning against the pattern the exploit demonstrates. Under the CyberJailbreak Severity scale that Anthropic, Amazon, Microsoft and Google co-published on July 4, GitLost would score CJS-2 as a jailbreak producing measurable exfiltration of confidential data through a supported product surface without customer misconfiguration beyond the vendor's own default template. A CJS-2 finding does not require CISA notification under the July 4 framework, but does trigger a coordinated-disclosure notice to affected model providers.