Daily2026-08-15
EconomicApplication

SpaceX closes $60 billion all-stock takeover of Cursor in largest coding-tools deal on record

SpaceX filed an 8-K at market close Friday confirming completion of the $60 billion all-stock acquisition of Anysphere, folding Cursor into the SpaceXAI unit alongside Grok Build, Grok Bot and Grok 4.6. Cursor shareholders receive SpaceX Class A stock at a 105 percent premium to the June IPO price. Morgan Stanley projects $2.5 billion of incremental 2026 revenue and $13 billion in 2027. Anthropic and OpenAI both filed European Commission comments urging deeper scrutiny of third-party model access foreclosure; the Commission cleared under Phase I without commitment offers.

Bloomberg / Yahoo Finance · 2026-08-14 · Source ↗

The 8-K hit the tape at 16:04 EDT Friday. SpaceX confirmed the close of its $60 billion all-stock acquisition of Anysphere, parent of Cursor, ending a takeover the two companies first sketched in April, formalized in June, and cleared through United States, European Union and United Kingdom regulators inside the June-to-August window. Cursor shareholders receive SpaceX Class A common stock at an implied equity value of $60 billion, a 105 percent premium to Cursor's June 12 IPO price of $29.3 billion. Anysphere becomes a wholly owned SpaceX subsidiary reporting into the SpaceXAI unit Elon Musk runs. The transaction is the largest coding-tools acquisition on record and dwarfs Microsoft's $7.5 billion GitHub deal in 2018.

The structure carries no cash. Every dollar of consideration flows through SpaceX equity. Cursor's engineering roster crosses over with the acquisition; CEO Michael Truell and cofounder Aman Sanger both take SVP roles under Musk with two-year non-competes and four-year equity vesting on their SpaceX Class A stock. Truell keeps the Cursor CEO title. Cursor's post on the official X account, published minutes after the 8-K, read: "Cursor is now part of SpaceX. Today, we have officially closed our acquisition. We will join the SpaceXAI team to help make Grok the world's most useful AI." Musk replied with a single word: "Welcome."

"Cursor is my favorite enterprise AI service." — Jensen Huang, Nvidia CEO, at last month's GTC keynote

The revenue math and the model concentration

Morgan Stanley published fresh models Friday afternoon showing Cursor at roughly $4 billion annual recurring revenue as of June 2026 and projecting a run-rate of $33 billion by 2030. The bank attributes $2.5 billion in incremental 2026 revenue to SpaceX and $13 billion in 2027, working from the assumption that SpaceX will bundle Cursor Ultra and Cursor Premium Teams into the SuperGrok Heavy subscription starting September 1. Cursor management said on the closing call that 64 percent of the Fortune 500 now use its tools, with 50,000-plus enterprises paying seat licenses. Cursor already contributed the training-data corpus and reinforcement-learning traces that produced Grok 4.5 in July, and Grok 4.6 shipped this week as the first Grok release integrated as a first-party model inside the Cursor IDE.

The engineering integration has been running for twelve weeks. Cursor engineers ran approximately 340 million tool-invocation trajectories through the Colossus supercomputer alongside the SpaceXAI post-training team, generating the reinforcement-learning-from-agent-execution corpus that lifted Grok 4.6 to 89.4 percent on Terminal-Bench 2.1, second only to Anthropic Opus 5. The Colossus cluster runs roughly a million H100-equivalent units of compute across two campuses in Memphis and Kingsburg, a scale that Cursor management said in the closing call would let the combined stack cut model training and post-training iteration time by roughly 60 percent versus Cursor's pre-acquisition setup.

"It's been a pleasure working with the SpaceXAI team on it." — Aman Sanger, Cursor cofounder, in an internal Loom video circulated to employees at 09:14 Pacific

The vendor foreclosure question the deal now poses

The competitive shape shifts. Anthropic ships Claude Code at v2.1.232 with Opus 5 as the default and auto mode as the default permission model as of August 14. OpenAI ships Codex CLI at 0.148.0-alpha.18. Meta ships Muse Code under a source-available license. DeepSeek shipped deepseek-harness under MIT the same week. SpaceXAI now controls Cursor plus Grok 4.6 plus the Grok Bot mobile client plus the Grok Build web agent. The five-vendor coding-agent race that opened in H1 2026 has consolidated into four vendors inside twelve hours, and three of the four hold both a frontier model and a coding-agent product in one shop.

Anthropic and OpenAI both filed European Commission comments urging deeper scrutiny of potential foreclosure of third-party model access inside Cursor, and both letters became public through the Commission's transparency portal Friday afternoon. Both labs have signaled they will pull first-party access to Cursor within the next twelve months, which would force SpaceX to route Cursor traffic exclusively through Grok models and threaten the enterprise-Fortune-500 base that Morgan Stanley's projections depend on. Prediction markets stayed skeptical of the strategic upside. Kalshi and Polymarket both show xAI at 13 percent probability of shipping the industry's best coding model by year-end, well behind Anthropic at 54 percent and OpenAI at 30 percent. The gap reflects both Grok's slower ramp on agentic-coding benchmarks and the fact that Cursor's differentiation historically came from routing across models rather than from model quality itself.

Regulatory calm surprised observers. The FTC and DOJ passed on second requests, and the European Commission cleared under Phase I without commitment offers. That contrasts with the Commission's more aggressive posture on last year's Microsoft-Inflection tie-up. SPCX closed down 1.8 percent on Friday at $131.62 on volume of 42.3 million shares, versus a 60-day average of 18.1 million. Cursor's revenue and headcount consolidate into SpaceX's financial reporting starting with the Q4 2026 quarter that begins October 1. Anthropic secondary marks stayed flat at $965 billion implied on Forge Global. OpenAI secondary marks moved to $860 billion, up from $852 billion.

Big picture

Friday's news cycle turned on consolidation and control. SpaceX closed the $60 billion Cursor acquisition at market close, the largest AI-tools transaction on record, folding a 64-percent-Fortune-500 coding surface into the same shop that ships Grok 4.6 and Colossus. Anthropic flipped Claude Code auto mode to the default across Pro, Max and Team plans and published a 1,053-tester study putting its classifier at 89 percent dangerous-command capture against 13.6 percent for human approvers. OpenAI opened the Ultrafast API preview running GPT-5.6 Sol on Cerebras wafer-scale silicon at 750 tokens per second, the first public non-Nvidia inference path for the flagship. Z.ai released GLM-5.3 to top CyberGym at 84.5 percent above Anthropic Mythos 5 and OpenAI GPT-5.6 Sol, then held weights back two weeks for safety hardening in the first Western-style staged open-weights release from a major Chinese lab.

The open-source and inference layers moved together. DeepSeek promoted V4-Pro-0813 to general availability with vendor-reported Terminal-Bench 2.1 at 87.9 and simultaneously raised API pricing up to 1,100 percent on cached input, then open-sourced its Harness agent runtime under MIT and hit 27,000 GitHub stars inside hours. Ollama v0.32.11 wired both DeepSeek Harness and Meta Muse Code into single-command launches at 01:22 UTC. llama.cpp shipped 11 builds between b10425 and b10435 in twelve hours covering virtual iGPU support, SYCL kernel fusion for q4_K quantization worth 2 to 12 percent throughput, and recurrent-state rollback for SSM operations. Claude Code v2.1.232 landed the auto-mode default flip alongside subagent forking, cross-session @-mentions and Opus 5 as the default Opus model.

Research, hardware and application news filled out the day. Hugging Face's Papers of the Day cluster was dominated by video world models: Alibaba's DreamX-Phi 1.0 took first on the WorldArena 2.0 robotics track with SE(3) attention encoding, Alaya-EVOKE ran an interactive world model at 23 fps on a single H200 through a camera-indexed memory bank, and HKU's PlayWorld benchmark tested nine world models across 171 interactive scenarios and found none clearing 50 percent on long-horizon aggregate scores. Shanghai AI Lab published Intern-S2-Preview at 397B with a 4B memory-decoder LoRA lifting Biology-Instructions by 3.4 points against a frozen backbone. OpenAI formalized Project Camellia in Effingham County, Georgia at 3.2 GW and at least $20 billion of private investment phased 2028 to 2032. Pure Data Centres committed EUR 1.5 billion for a 110 MW Finnish AI campus with a 550 MW expansion path. Uber and Pony.ai committed 2,000 Level 4 robotaxis to five European cities. Apple became the first foreign firm to clear CAC approval for a proprietary Chinese-market generative-AI model, trained with Alibaba Qwen team assistance. France's DGFiP confirmed the tax authority breach that exposed up to 2 million records via an MFA bypass detected in late June and disclosed only Friday after the attacker advertised the dataset on a Russian-language cybercrime forum.

More today

17 stories · 2026-08-14 — 2026-08-15
TechSafety

Z.ai ships GLM-5.3 with 84.5 percent CyberGym topping Mythos 5 and GPT-5.6 Sol, holds weights back two weeks

Decrypt / MarkTechPost · 2026-08-14 · Source ↗

GLM-5.3 keeps the 743B mixture-of-experts base from GLM-5.2 and lifts every capability from post-training alone. CyberGym at 84.5 percent edges Anthropic Mythos 5 at 83.8 and OpenAI GPT-5.6 Sol at 83.6. Z.ai delays open weights approximately two weeks for safety hardening, a first for a major Chinese lab.

Z.ai released GLM-5.3 on August 14 through its API, GLM Coding Plan and ZCode surfaces, framing the update as a demonstration that post-training scaling alone can push an open-weights model to frontier coding and cybersecurity benchmarks without a fresh pretraining run. The Beijing lab kept the 743-billion-parameter mixture-of-experts foundation from GLM-5.2 intact, layered additional reinforcement-learning and skill-training runs on top, and reports gains across every category the company tracks.

The coding numbers step across three benchmarks. Terminal-Bench 3.0 climbs to 28.3 from GLM-5.2's 4.6, a six-fold jump on the same base model. DeepSWE v1.1 lifts to 66.9 from 46.2. Agents' Last Exam CLI reaches 28.5 from 23.8. On GDPval-AA v2 the model posts 1,769 Elo, the highest score Z.ai has measured internally and above Anthropic Claude Fable 5 at 1,743. Terminal-Bench 2.1 sits at 88.2, level with GPT-5.6 Sol at 88.8. HumanEval-Plus lands at 91.4 percent, MBPP-Plus at 87.9, SWE-Bench Verified at 83.6 on Z.ai's internal run and 82.8 on the Papers With Code reproduction posted inside 24 hours.

Cybersecurity is where Z.ai says capability grew faster than the team expected. On CyberGym, which tests whether a model can identify and validate vulnerabilities from white-box source code, GLM-5.3 posts 84.5 percent, ahead of Claude Mythos 5 at 83.8 and OpenAI GPT-5.6 Sol at 83.6. On ExploitBench, the model doubles GLM-5.2's 24.4 to reach 54.4 percent, still behind Mythos 5 at 78.0. ExploitGym at the two-hour tier reaches 105 tasks against GLM-5.2's 29 and Mythos 5's 181; at six hours GLM-5.3 lands 130 tasks versus GLM-5.2's 39 and Mythos 5's 247.

"As we scaled post-training, cyber capability developed faster than we expected." — Z.ai, GLM-5.3 announcement

The staged open-weights choice

Z.ai disclosed a real-world test across 269 codebases where GLM-5.3 flagged 2,436 vulnerabilities, 1,097 rated medium-to-high severity, with an average bug age of 26.6 years. That figure carries the pull because it maps directly onto the class of aged open-source dependencies where enterprise defenders spend the most triage time. The company is holding weights back approximately two weeks pending safety evaluation and hardening, marking the first time a major Chinese lab has staged an open-weights release rather than shipping simultaneously with the API. Alibaba's Qwen 3.8-Max launched last week under a custom license with weights immediately live on Hugging Face and ModelScope. DeepSeek's V4-Pro-0813, released the same Friday morning, shipped API-only with no weights. Z.ai's plan to publish under Apache 2.0 in two weeks is the closest a Chinese lab has come to Western frontier-lab practice on responsible open-weights release.

"Scaling post-training is all we did for GLM-5.3." — Z.ai, model announcement

Pricing runs aggressively. Input tokens land at 2 yuan per million and output at 6 yuan per million, undercutting DeepSeek V4-Pro's new peak pricing by 78 percent on input and 78 percent on output. Context sits at 1 million tokens with 128K output, matching the GLM-5.2 window. Writer's Palmyra X6 flagship, released Thursday, is built as a post-training variation on GLM-5.2 and Writer sources said the company will evaluate GLM-5.3 as soon as weights become public, with a Palmyra X6.1 update on the newer base possible before year-end.

The CyberGym leadership carries international implications. CyberGym is a benchmark designed by UC Berkeley and Palo Alto Networks Unit 42 to test vulnerability discovery from source code, and it is one of four benchmarks the White House OSTP's June 2026 executive order on AI cyber capability references directly. A Chinese model taking the top slot on CyberGym gives the White House and BIS a fresh talking point in the ongoing debate over whether to expand Entity List coverage of Chinese AI labs. Zhipu received an Entity List designation in December 2024, but the designation did not stop the company from training and releasing GLM-5.3 on domestic Chinese Ascend hardware.

TechApplication

OpenAI opens Ultrafast API preview running GPT-5.6 Sol on Cerebras at 750 tokens per second

Help Net Security · 2026-08-14 · Source ↗

OpenAI opened a limited preview of an Ultrafast API tier Thursday running the flagship reasoning model on Cerebras wafer-scale accelerators at up to 750 output tokens per second and up to 14 times faster than the standard endpoint. First public non-Nvidia inference path for GPT-5.6 Sol.

OpenAI opened a limited preview of a new API tier called Ultrafast on August 14, running GPT-5.6 Sol on Cerebras wafer-scale accelerators at up to 750 output tokens per second and up to 14 times faster than the same model served through the standard endpoint. The tier represents the first public deployment of OpenAI's flagship reasoning model on hardware outside the Nvidia stack that dominates OpenAI's Azure and CoreWeave infrastructure.

The performance envelope matters because Sol is a reasoning model. Standard Sol produces long internal deliberation chains before returning tokens, and interactive latency has been the pain point for coding assistants, financial-analysis agents and support workflows. At 750 tokens per second, Sol holds the interactive speed floor that OpenAI has recommended for chat surfaces since GPT-4o and maintains it on a model that scores 61 on the Artificial Analysis Intelligence Index, tied with Grok 4.6 and second only to Claude Opus 5 and Fable 5.

"The increase in speed brought by Cerebras is impressive. It makes it practical for developers to work in a more focused and productive way." — John Crepezzi, Jane Street

The Nvidia dependency break

Cerebras confirmed the integration through its own press release the same window, billing Ultrafast under the ongoing $10 billion inference partnership the two companies signed earlier in 2026. Access is gated. OpenAI took Ultrafast live only for a slice of API customers who registered through a signup form and framed the tier as an interactive-workload extension rather than a general-purpose replacement for standard Sol. Pricing has not been fully published, but Ultrafast sits above the existing Fast Mode tier that runs at roughly 2.5x standard speed for roughly double the standard $5 in / $30 out per million tokens on Sol. OpenAI plans to expand access as Cerebras capacity comes online.

"GPT-5.6 Sol on Ultrafast is proof that speed and intelligence are no longer mutually exclusive." — Andrew Feldman, Cerebras CEO

Use cases OpenAI put in the launch materials cut across incident response with real-time log reading and trace analysis, financial market monitoring, e-commerce personalization and interactive research replacing overnight batch runs. Internally OpenAI is already routing Ultrafast to its own incident-response and research surfaces. Sub-second first-token latency lets a multi-step chain-of-thought agent complete its deliberation and tool-call cycle inside the timeout window of a typical HTTP client, opening application patterns that Sol on the standard tier could not sustain.

For OpenAI, plugging Sol into Cerebras hardware breaks a single-provider dependency chain and gives customers a benchmark against Nvidia H100 and Blackwell inference numbers on identical prompts. For Cerebras, hosting the flagship at production scale is the strongest possible validation ahead of the company's already-planned public offering. For enterprise buyers, the option to route latency-sensitive Sol traffic through a non-Nvidia path arrives at the moment the Nvidia-CoreWeave circular-financing dynamic is drawing scrutiny.

SafetyApplication

Anthropic flips Claude Code auto mode to default with classifier catching 89 percent of dangerous commands

Anthropic · 2026-08-14 · Source ↗

Auto mode became the default permission model for Claude Code on Pro, Max and Team plans on August 14. Anthropic's 1,053-tester controlled study measured 89 percent classifier catch of dangerous commands versus 13.6 percent for humans on first try, dropping to 5 percent after 50 approvals. Auto-mode users ship 25 percent more pull requests.

Anthropic's scheduled auto-mode default flip for Claude Code went live on August 14, changing the baseline permission behavior for all new sessions on Pro, Max and Team plans. In auto mode, Claude Code stops asking for human approval on every tool call and instead runs a classifier over each proposed action, halting only when the classifier determines the action is irreversible, destructive or reaches outside the user's environment. The change follows a week of tester messaging and one-time in-product opt-out prompts for users with pinned defaults.

The safety math Anthropic published to justify the flip is the load-bearing piece. In a controlled study of 1,053 external testers, auto mode's classifier caught 89 percent of dangerous commands on the first try. Human reviewers on the same trace set caught 13.6 percent, dropping to 5 percent after 50 approval prompts as fatigue set in. Real-world telemetry showed production-level harmful actions fired 2.6 times more often in manually approved sessions at 6.3 percent than in auto-mode sessions at 2.4 percent. In red-team testing after hardening, auto mode's miss rate improved from 12 percent to 7 percent. Anthropic reports zero successful prompt-injection attacks against Claude models under auto mode in the current battery.

"Employees view it as a weight off their shoulders. They don't have to monitor agents anymore." — Evan Magnussen, Garner Health Platform Engineering Manager

Rubber-stamp behavior collapsed the case for the old default

Human-approval behavior collapsed the case for the old default. Developers approved 97 percent of permission prompts, rejecting only 3 percent. On plans, users reject 39 percent of Claude's proposed plans but only 3 percent of individual permission requests once the plan is running. Humans review carefully at the plan stage and then rubber-stamp mechanically. Auto mode moves the guarantee from human vigilance to classifier vigilance and reserves human intervention for actions that leave the sandbox.

"I kicked off an agent at 10 p.m. and it kept running until 5 a.m., and it gave me three PRs." — Kai Zhou, Nuro Staff Software Engineer

Productivity data alongside the safety numbers pushes the release. Anthropic reports auto-mode users ship approximately 25 percent more pull requests. Adobe's Tomislav Reil said his team "built an agentic loop that rapidly accelerated our work." Gusto reports about 10 percent of sessions include classifier denials, indicating active protection is triggering regularly rather than sitting dormant. Classifier compute overhead is no longer billed to Pro, Max and Team plans starting the same week.

Auto mode stays opt-in on Claude Enterprise, the direct Claude API, Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry. Anthropic framed the enterprise carve-out as time for admins to review the classifier behavior against internal change-management policies before default flip. Users on Pro, Max or Team plans who had previously set a manual default receive a one-time prompt asking whether to switch; pinned defaults hold. The 89 versus 13.6 percent number is doing heavy lifting because it reframes the coding-agent oversight debate away from human-in-the-loop dogma and toward whether the classifier is actually right.

TechEconomic

DeepSeek V4-Pro-0813 hits GA with 87.9 Terminal-Bench, lifts pricing up to 1,100 percent on cached input

The Decoder / Caixin Global · 2026-08-14 · Source ↗

DeepSeek promoted V4-Pro to general availability Friday with vendor-reported Terminal-Bench 2.1 at 87.9, near Claude Opus 4.8 and Fable 5. API pricing shifts to peak/off-peak billing effective August 16 at 16:00 UTC, with cached-input tokens up 1,100 percent to 0.3 yuan per million.

DeepSeek moved V4-Pro from the April 24 preview to general availability on August 13 US time, with the update landing in Western press on August 14. The API endpoint deepseek-v4-pro now routes to V4-Pro-0813 with no code change required. The model keeps the 1M-token context window and the 671B total parameter count from the preview but ships materially better agent behavior according to DeepSeek's own numbers.

The vendor-reported jumps against the April preview are the headline. Terminal-Bench 2.1 climbed from 72.1 to 87.9, near Claude Opus 4.8 at 85.0 and Fable 5 at 88.0. CyberGym went from 52.7 to 83.3, edging past Opus 4.8's 78.3. DeepSWE rose from 12.8 to 62.7. AutomationBench went 12.8 to 31.8. DSBench-Hard climbed 31.1 to 67.2. HLE without tools measured 42.7, with tools 60.0. NL2Repo landed at 61.5. Toolathlon-Verified came in at 74.1. Artificial Analysis Intelligence Index moved from 45 in preview to 53 in the GA build, tied with GLM-5.2 and behind Claude Opus 5 at 63 and Kimi K3 at 60. Papers With Code opened a reproduction request the same day since DeepSeek published methodology inside its newly open-sourced Harness repository.

"V4-Pro-0813 substantially improves its capabilities for AI agents." — DeepSeek, model announcement

The pricing lever DeepSeek pulled

Pricing is the counter-narrative. V4-Pro-0813 sits at $1.32 per million input tokens and $3.96 per million output tokens versus V4-Flash at $0.14 input and $0.28 output. That is 9x on input and 14x on output relative to Flash. Effective August 16 at 16:00 UTC DeepSeek switches to a peak/off-peak schedule with off-peak set at half of peak. Peak input runs $1.32, off-peak $0.66. Peak output runs $3.96, off-peak $1.98. Cache hits jumped hardest: $0.022 peak and $0.044 off-peak, up from $0.003625 in the preview era, a 6-12x price increase depending on tier. Caixin Global put the headline cached-input increase at 1,100 percent from 0.025 yuan to 0.3 yuan per million tokens.

"Off-peak prices set at half of the peak-hour prices." — DeepSeek pricing note

The pricing shift lands 24 hours after DeepSeek reportedly attempted to promote V4-Pro-0813 to general availability Tuesday and then pulled the announcement inside 24 hours as Artificial Analysis's Intelligence Index scoring landed at 53 rather than the expected 55-plus. Friday's public release marks the second attempt, this time with the pricing restructure included in the initial announcement rather than added later. Twelve of the top twenty largest Chinese-domiciled V4-Flash API accounts, contacted by Caixin Friday afternoon, said they will scale down commitments and route higher-value workloads to Alibaba Qwen 3.8-Max instead. Bernstein published a Friday-afternoon note projecting a 30-to-40 percent volume decline in DeepSeek's Chinese API business over the next quarter as workloads route to Qwen and GLM-5.3.

Even at the new rates DeepSeek V4-Pro sits below OpenAI GPT-5.6 Sol at $5 input and $30 output per million tokens and Claude Fable 5 at $10 input and $50 output. The gap narrows without closing. Nvidia gained 1.2 percent Friday on the pricing news, as buy-side analysts read the shift as evidence that even DeepSeek's cost curve had reached the level where per-token pricing must reflect the economics of high-end GPUs at scale. The pricing decision reduces the "DeepSeek deflation" narrative that had weighed on Nvidia data-center revenue outlook.

TechApplication

DeepSeek open-sources Harness agent framework under MIT, 27,000 GitHub stars inside hours

BigDATAwire (HPCwire) · 2026-08-14 · Source ↗

DeepSeek published deepseek-harness on GitHub under MIT as a developer preview built on the Cordis plugin kernel. Every capability — models, tools, sandboxes, sessions, loops, orchestration, UI — is a swappable plugin. The repo crossed 27,000 stars and 2,000 forks the launch day.

DeepSeek shipped deepseek-harness v0.1 as a developer preview at github.com/deepseek-ai/deepseek-harness under the MIT license, opening the runtime layer that ties DeepSeek's V4-Pro and V4-Flash models to autonomous coding-agent behavior. The repository crossed 27,000 GitHub stars and roughly 2,000 forks within hours of the launch tweet from the official DeepSeek account, according to coverage from VentureBeat, TechNode and BigDATAwire on August 14.

The design pitch is one line on the README: everything is a plugin. Models, tools, skills, sessions, sandboxes, filesystems, loops, scheduling and the web UI all sit behind the pre-existing Cordis plugin kernel that DeepSeek's UI team has used internally since 2023. Sessions run through an append-only event log, so any session can be resumed, forked, searched or replayed. That matches the architecture Anthropic added to Claude Code's session store in v2.1 but does it under a permissive license and outside a vendor console.

"Everything is a plugin. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, UI." — deepseek-harness README

Four preset runtime modes and the SpaceX-Cursor timing

The v0.1.0-rc.5 release ships four preset runtime modes. Standard mode is the daily-driver agent with file editing, shell access, web search, skills and planning tools. Code mode replaces the wire-level tool set with a single run_code call and provides TypeScript SDK bindings for agent authors who want the model to write executor code rather than pick from a menu. Minimal mode strips the toolset to two calls, persistent bash and str_replace_editor, and documents DeepSeek's V4-Pro-0813 benchmark methodology in the repo's BENCHMARK.md footnotes. Creator mode adds runtime inspection endpoints for developers writing new plugins.

"We're opening it up to developers building agent harnesses worldwide and open-sourcing the codebase in MIT license." — DeepSeek official X account

Positioning is unsubtle. The project frames itself directly against Anthropic's Claude Code and Claude Cowork, both of which run under Anthropic's proprietary agent runtime with no equivalent extension surface. Meta's Muse Code, released August 10 under Apache 2.0, targets the same coding-agent niche but ships as a monolith without the plugin kernel. OpenAI's Codex CLI is Apache-2.0 but its plugin surface is limited to MCP servers and shell tools. Cui Tianyi, who joined DeepSeek from Jane Street in March 2026 to build the harness team, leads the project.

Two integrations landed inside 24 hours. Ollama v0.32.11 added an ollama launch dsh command that boots deepseek-harness against a locally served model, and the Codex CLI 0.148.0-alpha.18 build gained a --harness deepseek flag that routes agent loops through Harness rather than Codex's built-in runner. The release also gives outside developers a reason to keep using DeepSeek's models directly rather than through hosted alternatives, offsetting the pricing hike for V4-Pro. The README warns "THERE WILL BE COMPATIBILITY-BREAKING CHANGES" during the preview phase, with a v1.0 stable target for Q4 2026. The MIT license and public BENCHMARK.md configuration together open a clean replication path for DeepSeek's disputed benchmark claims.

TechSafety

Claude Code v2.1.232 ships auto-mode flip, subagent forking, cross-session @-mentions, Opus 5 default

Anthropic Claude Code Changelog · 2026-08-14 · Source ↗

Anthropic promoted Claude Code v2.1.232 alongside the auto-mode default flip. Subagent forking inherits full conversation and prompt cache; typing @ in a prompt pops a live-session picker that routes through SendMessage; Opus 5 is the new default Opus model with 1M-token context on Max, Team and Enterprise.

Anthropic promoted Claude Code v2.1.232 on August 14 alongside the flip of auto mode to the default permission model for new sessions on Pro, Max and Team plans. Enterprise and cloud deployments remain opt-in. The version bump ships four functional additions and a wide fix set for Remote Control, permissions, networking and session stability.

The subagent forking change is technical. A subagent_type: "fork" subagent now inherits the full conversation and prompt cache from the parent session by default, matching how Anthropic's own internal agents behave rather than the previous fresh-context spawn. Non-teammate agent spawns in interactive sessions run in the background by default now, which means the parent session keeps the terminal and the agent posts progress via SendMessage.

"Subagent forking is now on by default." — Claude Code v2.1.232 changelog

Cross-session messaging, Opus 5 default, self-hosted runners

Cross-session messaging is new. Typing @ in a prompt now pops a live-session picker; Claude uses SendMessage to reach the mentioned session directly. Bare-name delivery works when exactly one live session matches; ambiguous names trigger a confirmation dialog. Session names must be unique across a machine, and duplicate names get name-word-word variants automatically. Claude Opus 5 is the new default Opus model in Claude Code. On the Anthropic API and on Max, Team and Enterprise plans, Opus 5 runs with a 1M-token context window. Fable 5 is available as an advisor model through /model fable for organizations with Fable access.

"Type @ in the prompt to mention another Claude session by name." — Claude Code v2.1.232 changelog

The changelog covers additional platform work. GitLab and marketplace support tightened: bare GitLab URLs now clone like GitHub URLs; secret redaction extended to multiple additional token families. Archive plugin sources can install from a zip over HTTPS without git or npm, with optional SHA-256 pinning. Self-hosted-runner support turns customer-owned machines or containers into a place Claude Code web, mobile and desktop sessions can run on Team and Enterprise plans, a first for the product. Settings aliases additionalMarketplaces and allowedMarketplaces replace previous naming conventions with a compatibility shim.

Gateway validation moved earlier in the boot sequence. The desktop overlay now accepts all released settings, and invalid entries fail at process startup rather than at first tool call. This addresses a class of MCP OAuth failures that plagued Claude Code v2.1.229 through v2.1.231, including the Slack sign-in regression Anthropic hot-fixed inside 24 hours on August 13. Named fixes include permission bypasses on Windows and PowerShell where session context was not propagating to spawned subprocesses, MCP connection hangs when the remote server returned a chunked-encoded 429, Remote Control session reattachment failures after network changes, TUI paste handling of large clipboard payloads, and a slash-command menu positioning bug on high-DPI displays.

The v2.1.232 build closes a two-week release cadence that produced eight patch versions between v2.1.225 and v2.1.232, more than any prior fortnight in Claude Code's release history. Anthropic's release notes attribute the cadence to the auto-mode flip readiness gates.

TechApplication

Ollama v0.32.11 wires DeepSeek Harness and Meta Muse Code into single-command launches

Ollama GitHub Releases · 2026-08-14 · Source ↗

Ollama v0.32.11 landed at 01:22 UTC on August 14 adding ollama launch dsh for DeepSeek's fresh Harness, ollama launch muse for Meta's Muse Code CLI, and web-search support in the OpenAI-compatible Responses API. First-class launchers for competitor agent frameworks.

Ollama shipped v0.32.11 at 01:22 UTC on August 14, ten hours after DeepSeek open-sourced its Harness agent framework and four days after Meta released Muse Code. The release adds two launch commands that boot the third-party harnesses against a locally served Ollama model, plus a web-search capability inside the OpenAI-compatible Responses API endpoint.

The DeepSeek Harness integration is the headline. Running ollama launch dsh at the terminal now pulls the deepseek-harness developer preview from npm, wires the Ollama-served model as the harness's default endpoint, and drops a web UI at 127.0.0.1:3080. Users can choose between DeepSeek's V4-Pro or V4-Flash if they have API keys, but the primary use case is running Harness against local models such as Qwen 3.5-Coder, DeepSeek V3.2 or Meta Muse Glimmer 30B. This is the first time an inference server has shipped a first-class launcher for a competitor's agent framework, and it matches what Ollama did with LangGraph in v0.31.

"ollama launch dsh integrates with DeepSeek's open-source agent harness." — Ollama v0.32.11 release notes

The Muse Code path and the Responses API extension

The Muse Code integration takes the same shape. ollama launch muse pulls Meta's agentic coding CLI, released August 10 under a source-available license, and points it at the local Ollama daemon. Muse Code is optimized for Muse Glimmer 30B and Muse Spark 1.2, both of which Ollama already supports natively across NVIDIA, AMD and Apple Silicon backends. The Responses API extension adds a web_search tool that mirrors OpenAI's server-side search behavior. Ollama-hosted models can now call web search as a native tool from within a Responses API session without a separate LangChain or LlamaIndex adapter. The implementation uses SerpAPI as the default backend with Brave Search and DuckDuckGo as alternatives selectable at daemon boot.

Muse Glimmer template improvements landed in the same version, tightening how Ollama's Jinja renderer handles the reasoning-block scaffolding Meta's model expects. This closes an issue Meta engineers reported August 11 where Muse Glimmer's reasoning traces truncated at 8K tokens rather than the intended 32K. The v0.32 series has now shipped 12 releases since v0.32.0 debuted on July 11, and the median gap between v0.32 releases is 3.1 days, compared to 5.7 days across the v0.31 series.

"ollama launch muse provides access to Meta's agentic coding CLI." — Ollama v0.32.11 release notes

The launch commands collapse harness setup into a single line, closing a gap that had held back local-first agent adoption. Prior to v0.32.11 running either harness against a local Ollama model required manual configuration of the model endpoint URL, tokenizer and prompt-template registry. The pattern matches how Ollama handles model installation with ollama run and how Claude Code handles workspace setup with claude init.

Tech

llama.cpp ships eleven builds in twelve hours: virtual iGPU, chat-template reasoning effort, SYCL kernel fusion

llama.cpp GitHub Releases · 2026-08-14 · Source ↗

llama.cpp landed builds b10425 through b10435 across August 14 covering virtual iGPU support for hypervisor-guest inference, chat-template reasoning_effort passthrough matching OpenAI, SYCL FFN kernel fusion for q4_K worth 2–12 percent throughput, and recurrent-state rollback for SSM operations.

llama.cpp shipped eleven builds on August 14, from b10425 through b10435, one of the densest release days in the project's history. The batch spans hardware-backend work on Intel SYCL and virtual iGPU support, format work on chat-template reasoning-effort passthrough, correctness fixes on server slot metrics and a jinja template regression fix at the top of the stack.

The virtual iGPU device support in b10430 is the most substantive addition. Modern integrated GPUs on Intel Meteor Lake and AMD Strix Point expose device-descriptor variants beyond the standard OpenCL and Vulkan enumerations, and b10430 adds a fallback path that lets llama.cpp target the virtualized device presented by hypervisors when the guest OS does not see the physical iGPU directly. This is a build-out for the growing set of AI workloads running inside virtualized Windows containers on Linux hosts, and it lands one week after Microsoft added similar support to DirectML.

"b10434: Chat template reasoning effort support." — llama.cpp release notes

SYCL, SSM rollback, jinja fix

Build b10434 adds a reasoning_effort parameter to Jinja chat templates for OpenAI Chat Completions compatibility. The parameter accepts low, medium and high values matching OpenAI's Responses API and DeepSeek V4-Pro-0813's three-tier reasoning effort selector. Templates that inspect reasoning_effort can now emit different system-prompt scaffolding, thinking-block markers or tool-call schemas depending on the caller-selected tier without requiring template branching in the client.

The SYCL work runs across two commits. Build b10427 fuses the FFN kernel for q4_K quantization on Intel Data Center Max, Flex and Arc GPUs, worth 2 to 12 percent throughput gains on the reference Llama 3.1 8B and Qwen 3.5 14B workloads. Build b10425 fuses the gated-delta-net state write path with a further roughly 1 percent gain on models that use SSM state variables. Together the two commits close the SYCL backend's throughput gap against the CUDA backend on the q4_K quantization path from 18 percent to roughly 4 percent on Arc B580.

"b10431: Recurrent state rollback for SSM operations." — llama.cpp release notes

Build b10431 adds recurrent-state rollback for SSM operations with CPU and CUDA backend support and a fallback mechanism for unsupported backends. This lets models with recurrent components including Mamba, RWKV and hybrid SSM-attention architectures roll back internal state during speculative decoding without recomputing from the last checkpoint, previous fallback cost 30 to 50 percent of the speculative-decode benefit on those models. Build b10429 opens the metrics and slots endpoints of the llama.cpp server for read access during decode operations. Build b10433 syncs the vendored GGML library, pulling in three new SIMD kernels for AVX-512 BF16 and one bug fix for the ARM SVE dot-product path on Neoverse N2. Build b10435, the top of stack at end of day, fixes a quadratic-cost regression in the jinja gather_string_parts helper introduced by b10434's reasoning-effort work.

All eleven builds publish pre-built binaries for macOS Intel, macOS ARM64, Ubuntu x64 with and without OpenVINO 2026.2.1, Ubuntu ARM64, Ubuntu SYCL fp16 x64, Ubuntu s390x, Windows x64 with CUDA 12 and 13, Android arm64-v8a and source archives.

TechTrend

Alibaba DreamX-Phi 1.0 tops Hugging Face robotics chart with SE(3) attention encoding

arXiv / Hugging Face Papers · 2026-08-14 · Source ↗

Alibaba's DreamX team publishes an action-conditioned video world model for robotic manipulation that took first place on Track 1 and second on Track 2 of WorldArena 2.0. The 79-upvote paper injects per-arm SE(3) transformations into attention through PRoPE-style geometric encoding, distilled to run at 23 fps.

Alibaba's DreamX Team pushed DreamX-Phi 1.0 to arXiv (2608.13489) on August 13 UTC, and Hugging Face pinned it at position three of the Papers of the Day cluster on August 14 with 79 upvotes as of 18:00 UTC. The model predicts future observations given an observed frame, a language instruction and a prescribed action sequence comprising end-effector poses and gripper states, positioning it as a foundational asset for the video-world-model class of robotic manipulation stacks that Google DeepMind, Physical Intelligence and NVIDIA GR00T have all pursued through the first half of the year.

The architecture stacks four components that the paper claims each address a specific failure mode of prior action-conditioned video world models. PRoPE-style geometric encoding injects per-arm SE(3) transformations directly into the attention mechanism, giving the model an explicit geometric anchor for each robot arm rather than relying on the pixel encoder to induce the frame from raw video. A dedicated depth branch handles scene-level geometry so that near-field and far-field relationships stay stable across the prediction horizon. SAM3 masks paired with a frozen V-JEPA teacher hold object identity through occlusion during grasping, addressing the long-standing failure of video world models where a manipulator hand hides a target and the model paints in a different object on the other side. Distribution-matching distillation compresses multi-step generation into a few-step student, cutting inference latency to a range the paper says is compatible with closed-loop policy roll-outs.

"PRoPE-style geometric encoding injecting per-arm SE(3) transformations into attention mechanisms." — DreamX-Phi 1.0 paper

WorldArena 2.0 leaderboard and license positioning

The authors evaluated the system on WorldArena 2.0, which introduced tightened action-following metrics after the original benchmark got saturated in late 2025. DreamX-Phi 1.0 took first on Track 1, which scores unconditional-generation fidelity against ground-truth demonstrations, and second on Track 2, which scores action-conditioned open-loop rollouts against a hidden evaluation set. The repository, published at AMAP-ML/DreamX-Phi on GitHub, ships checkpoints for the base video generator, the depth branch, the distilled student and integration hooks for Isaac Sim and Meta Habitat.

"SAM3 masks with frozen V-JEPA teacher for object consistency during grasping." — DreamX-Phi 1.0 paper

The paper credits Rui Chen, Xiangxiang Chu, Geng Li, Jifan Li, Qingfeng Shi, Datao Tang, Jing Tang, Jun Wang and Pengfei Zhang as named contributors, with the broader DreamX Team listed as corresponding author. The DreamX group sits inside Alibaba's AMAP autonomous mobility organization, which historically focused on mapping and delivery-robot fleets before pivoting toward general manipulation with the release of the DreamX-Air aerial simulator in April. Alibaba's decision to publish weights alongside the paper is consistent with the pattern that started with the Qwen 3.8-Max release earlier in the week, when Alibaba moved to a custom Kimi-style license with revenue-share and geographic-authorization clauses. DreamX-Phi 1.0 ships under Apache 2.0 according to the repository, keeping the manipulation stack outside the tighter licensing regime that now covers Alibaba's frontier language models. The paper does not report an inference cost figure per rollout, and independent teams looking to reproduce the manipulation numbers will need to hold the model on a single H200 through at least the distilled configuration, which the paper describes as running at approximately 23 frames per second.

TechTrend

Alaya-EVOKE bounds interactive world-model memory with camera-indexed state bank, runs 23 fps on single H200

arXiv / Hugging Face Papers · 2026-08-14 · Source ↗

A six-author group led by Yuanyang Yin publishes Alaya-EVOKE, an interactive video world model that stores scene geometry in an external camera-indexed bank and generates each 1.5-second chunk in 2.11 seconds on a single H200 GPU at 384×640 resolution. Second on HF Papers at 82 upvotes.

The Alaya-EVOKE paper (arXiv 2608.13546) landed at 82 upvotes and second position on Hugging Face Papers of the Day on August 14 UTC, one slot ahead of Alibaba's DreamX-Phi robotics model. Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang and Feng Zhao filed the paper under Computer Vision and Pattern Recognition (cs.CV) on August 13, and the accompanying inference stack ships as a supplementary release from the Alaya group at Shanghai AI Laboratory.

The subtitle "From Linear-Scaling Supervision to Endless World" telegraphs the paper's central claim: interactive world models today choke on memory as session length grows because both the model and its supervision cost scale linearly with visited frames. Alaya-EVOKE separates the geometry substrate from the generative denoiser by maintaining scene geometry in an external camera-indexed world state bank, from which only view-relevant information is retrieved at each step. The retrieved subset feeds into a bounded denoiser context, freezing peak memory even as the agent explores new rooms or environments across arbitrarily long horizons.

"Scene geometry in an external, camera-indexed world state bank, from which only view-relevant information is retrieved." — Alaya-EVOKE paper

Sparse-attention teacher and three-step student

The teacher design pairs the memory retrieval with a supervision signal that also stops scaling linearly. Sparse attention combines chunk-wise grouping, retrieval of selected distant frames and a linear-attention global state, letting the teacher supervise very long roll-outs at bounded compute per token. The framework then transfers to a three-step student model, dropping classifier-free guidance and cutting the wall-clock cost of interactive inference.

"Each 1.5-second chunk generates in 2.11 seconds on a single H200 GPU at 384×640 resolution." — Alaya-EVOKE paper

Reported throughput numbers are unusually explicit for a world-model paper. Each 1.5-second chunk generates in 2.11 seconds on a single H200 GPU at 384×640 resolution, and the open-ended generation loop clocks in at approximately 23 frames per second sustained. The paper posts state-of-the-art on WBench and stays within one point of the leader on VBench-Long and VBench-2.0, the two benchmarks the world-model community leaned on through the first half of 2026.

The camera-indexed state bank marks a departure from the direction that OpenAI's Sora-2 preview and Google DeepMind's Genie-4 have taken, both of which lean on ever-larger context windows inside a monolithic transformer stack. Alaya-EVOKE's authors bet that the correct primitive is retrieval against an explicit geometric index rather than dense attention across all past frames, the same bet that Meta's V-JEPA and the RT-World stack made earlier this quarter. The paper does not report closed-loop policy scores, so the robotics community will need to run its own trials before adopting Alaya-EVOKE as a substrate for policy learning.

TechSafety

HKU PlayWorld benchmark finds nine world models fail long-horizon interactive objectives

arXiv / Hugging Face Papers · 2026-08-14 · Source ↗

The University of Hong Kong publishes PlayWorld, a 171-scenario benchmark that uses multi-modal AI agents as interactive players against video world models. Testing across nine advanced models finds current systems unreliable on spatial consistency and persistent state evolution, none clearing 50 percent on the long-horizon aggregate.

PlayWorld (arXiv 2608.13552) landed on Hugging Face at 48 upvotes on August 14 UTC, filed by a twelve-author group led by Kaixin Ding at the University of Hong Kong with contributors spread across Kuaishou's Kling video generation team and Hengshuang Zhao's HKU lab. The paper opens a second front in world-model evaluation by dropping the fixed-action-sequence paradigm and letting a multi-modal agent choose actions online against the world model under evaluation, then scoring the resulting trajectory against a hidden objective.

The benchmark carries 171 distinct evaluation scenarios, each with a defined objective ranging from single-room navigation with hidden targets to multi-step manipulation puzzles that require the model to remember object placements from earlier frames. Each scenario is scored across four primary assessment areas: geometry consistency, interaction fidelity, out-of-sight evolution and insight evolution, together with basic video quality and controllability metrics that the older VBench line already covered.

"Agent players to interact with world models toward specified long-horizon objectives." — PlayWorld paper

The agent-player design and the uniformly negative result

The core innovation is the agent player. Rather than feeding the world model a curated ground-truth action sequence and asking whether the model reproduces the ground-truth video, PlayWorld hands a multi-modal AI agent the objective and the current frame, lets it pick the next action, and evaluates the entire on-policy roll-out. This surfaces failure modes that fixed-sequence evaluation misses, particularly the class of drift where a model produces plausible individual frames but breaks the persistent state across a long session.

"Current models remain unreliable on long-horizon interactive objectives, particularly in maintaining spatial consistency and persistent state evolution." — PlayWorld paper

Testing across nine advanced world models produced a uniformly negative result. The paper does not name the nine models in the abstract, but the benchmark repository at kxding.github.io/project/PlayWorld/ names Sora-2, Genie-4, Kling 2.5, DreamX-Phi 1.0, Alaya-EVOKE, RT-World 3, V-JEPA-3, WHAM-2 and Gen-4 as the tested set. None of the tested models clear 50 percent on the long-horizon aggregate. The benchmark's construction is deliberately hostile to the retrieval-based memory approaches that Alaya-EVOKE and RT-World 3 pushed earlier this month. Out-of-sight evolution scores whether the world updates correctly when the objects that would be updated are outside the current viewport, which forces the model to maintain state even for content it cannot see. Insight evolution scores whether the model correctly re-integrates the returning viewport with its remembered state.

The 171 scenarios ship as reproducible seeds against the Kling video-generation platform and Kuaishou's evaluation cluster, which lowers the bar for third-party reproductions. Hengshuang Zhao's lab at HKU has published four benchmarks in 2026 already, all under permissive licenses. The benchmark's ceiling is intentionally out of reach for current models, matching the pattern GAIA-Long and Terminal-Bench 2.1 established for language agents earlier this year.

TechApplication

Shanghai AI Lab Intern-S2-Preview at 397B posts scientific-agent gains via 4B memory decoder

arXiv / Hugging Face Papers · 2026-08-14 · Source ↗

A 125-author Shanghai AI Lab team publishes Intern-S2-Preview, a 397-billion-parameter scientific foundation model paired with a 4-billion-parameter Intern-MemDec memory decoder that lifts Biology-Instructions from 56.92 to 60.32 without touching the frozen backbone.

Intern-S2-Preview (arXiv 2608.13505) went live at 42 upvotes on Hugging Face Papers of the Day on August 14 UTC, submitted August 13 by lead author Lei Bai with 124 co-authors under the Intern Large Models umbrella at Shanghai AI Laboratory. The release covers two model artifacts: Intern-S2-Preview-397B, a scientific foundation model, and Intern-MemDec-4B, a memory-augmented decoder that specializes the backbone without changing any of its frozen parameters.

The training recipe stacks four supervision phases that the paper claims each address a distinct failure mode of prior scientific foundation models. Supervised fine-tuning covers 340 million scientific instruction-following pairs across biology, chemistry, materials science, astronomy and mathematics. Scalable multi-task reinforcement learning covers 40 tasks that the paper says are chosen to force the model to hold multiple problem contexts simultaneously. Black- and white-box agentic RL closes the loop against verified scientific tools ranging from Chembl to RDKit to the OpenAstro sky simulation. On-policy distillation transfers capabilities into smaller sibling models without the reward-hacking pathologies that offline distillation induced in earlier iterations.

"Biology-Instructions score improved from 56.92 to 60.32 using the memory decoder." — Intern-S2-Preview paper

The memory-decoder pattern versus full-parameter fine-tuning

The paper reports competitive or leading results across scientific, multimodal, agentic and general-purpose benchmarks and enhanced performance on SciTS, a time-series understanding-and-forecasting suite that the Intern group released in June. The headline number is the Biology-Instructions score lift from 56.92 to 60.32 driven purely by the Intern-MemDec-4B decoder, which the paper argues shows that memory decoders offer a cheaper path to domain specialization than full-parameter fine-tuning of the 397B backbone.

"Memory-augmented extension enabling specialization without modifying the frozen backbone." — Intern-S2-Preview paper

The 397B parameter count places Intern-S2-Preview between Meta's Muse Glimmer at 30B and Alibaba's Qwen 3.8-Max at 2.4T total with 95B active, and closer in scale to Cohere Command R+ 2026 (280B) than to the frontier mixture-of-experts models. The architecture is described as dense rather than sparse in the paper, which departs from the MoE norm that DeepSeek V4-Pro and Kimi K3.1 established earlier this quarter. The memory decoder receives the paper's second-most attention. Intern-MemDec-4B is a small decoder that reads from and writes to a memory pool during inference, letting the frozen 397B backbone stay static while the memory-decoder LoRA learns per-domain specializations. The Biology-Instructions gain of 3.4 points at only 4B of trainable parameters is the paper's core evidence that this pattern generalizes.

The paper's 125-author list makes it one of the larger foundation-model publications of 2026, larger than the OLMo 3 preview at 82 authors and smaller than the original Kimi K3 paper at 173. The size reflects the tool-integration scope: the paper lists 40-plus external scientific tools that the model calls, and each tool required a domain team to build the adapter. Intern-S2-Preview does not report an intelligence-index score, so direct comparisons against GPT-5.6 Sol, Claude Opus 5 or Grok 4.6 are not possible from the paper. The scientific-benchmark strength suggests the model is targeting the same specialist niche that Anthropic's Claude for Legal and OpenAI's Astra Preparedness team have started to work on.

EconomicPolicy

OpenAI details Project Camellia 3.2 GW Georgia campus at $20 billion private investment

OpenAI · 2026-08-14 · Source ↗

OpenAI publishes the community brief on Project Camellia, a 3.2 GW data-center campus in Effingham County, Georgia representing at least $20 billion in private investment. Georgia Power delivers in phases 2028 to 2032. OpenAI pays all infrastructure and electric-service cost.

OpenAI published the community brief on Project Camellia on August 14 UTC, formally scoping the previously reported Effingham County site as a 3.2 GW data-center campus representing at least $20 billion in private investment. Effingham County sits roughly 45 minutes northwest of Savannah, and the campus will comprise four buildings across approximately 101 acres inside the 2,600-acre Savannah Gateway Industrial Hub, a site already zoned for industrial use.

Georgia Power will deliver up to 3.2 gigawatts of electricity in phases between 2028 and 2032. The OpenAI post commits that OpenAI will pay all infrastructure and electric-service costs, meaning residential customers will not subsidize the facility. The commitment matches the framework the company established with the Abilene, Texas campus and with the Stargate-linked builds in New Mexico and Wisconsin, and is a direct response to the utility-rate-shift complaints that surfaced when Meta's Louisiana campus and Amazon's Virginia buildouts pushed household electric bills higher in 2024 and 2025.

"Georgia families will not subsidize this project. OpenAI will pay all infrastructure and electric service." — OpenAI Newsroom

Closed-loop cooling and community terms

The community-benefits package covers $80 million allocated toward local education, public safety, healthcare and small-business resources over the initial ten-year operating window, per the OpenAI post. Effingham County officials estimate that OpenAI will become the county's largest taxpayer once Phase 1 is operational, generating hundreds of millions of dollars in state and local tax revenue over the campus's lifetime. Environmental design centers on closed-loop cooling that recirculates water, similar to a car radiator. Ongoing water use, per the OpenAI post, is expected to be comparable to a similar-size office building once fully built. The design choice sidesteps the aquifer-drawdown criticism that Google faced in Uruguay and Chile, and that Meta faced at its Prineville, Oregon campus, though the trade-off is higher upfront capital and higher electric load for the cooling loop.

"3.2 gigawatts of electricity, delivered in phases between 2028 and 2032." — OpenAI community brief

The 3.2 GW nameplate places Project Camellia among the largest single AI data-center campuses under development. For scale, CoreWeave's CW5 Tulsa campus is spec'd at 1.2 GW, Meta's Hyperion Louisiana campus at 2.5 GW, and Amazon's Pecos Texas campus that Public Citizen sued over in early August targets 7.65 GW. The Camellia phasing to 2032 suggests OpenAI is treating the site as the anchor for a second-generation Stargate cluster, complementing the Abilene campus which will hit first-power on the initial 250 MW block later this quarter. Georgia Power's Vogtle nuclear expansion added two AP1000 reactors in 2023 and 2024, and the Camellia phasing lines up with the second wave of Vogtle-linked transmission expansion that the Georgia Public Service Commission approved in March.

The community brief formalizes the reveal OpenAI staff socialized with local media in July, ahead of the site rezoning vote that Effingham County commissioners approved 4-1 on July 22. The Current Georgia and the Effingham Herald had reported the $20 billion capital figure at the time, but the OpenAI post is the first primary-source confirmation of the 3.2 GW power target, the 2028-to-2032 phasing schedule and the $80 million community-benefits figure.

EconomicTech

Pure Data Centres commits EUR 1.5B for 110 MW Finnish AI campus with 550 MW expansion path

Data Center Knowledge · 2026-08-14 · Source ↗

Pure Data Centres Group breaks ground on a 110 MW Phase 1 campus in Seinäjoki, Finland at EUR 1.5 billion, with an expansion path to 550-plus MW at a total EUR 7.5 billion. Racks designed for 130 kW direct-to-chip liquid cooling versus 60 kW at older Nordic sites.

Pure Data Centres Group disclosed the Seinäjoki, Finland campus on August 14 UTC via the Data Center Knowledge round-up, committing EUR 1.5 billion (roughly $1.7 billion) to a 110 MW Phase 1 build with an expansion path to more than 550 MW at a total EUR 7.5 billion (roughly $8.6 billion). The project takes the top slot in Europe's AI-focused infrastructure pipeline for 2026 and marks Pure Data Centres' third Nordic campus after the Espoo and Stavanger builds.

The Seinäjoki site sits in western Finland along the Vaasa-Tampere transmission corridor, giving the campus access to the Finnish grid's high renewable share. Fingrid, the Finnish transmission-system operator, reports that 92 percent of Finnish grid electricity in the first half of 2026 came from renewable or nuclear sources, driven by the completion of the Olkiluoto 3 reactor in 2023 and the ongoing offshore-wind build-out in the Gulf of Bothnia. Pure Data Centres describes the campus as renewable-powered without naming specific PPA counterparties in the initial disclosure.

"One of Europe's largest AI infrastructure projects." — Data Center Knowledge, Aug 14

Rack density, permitting, and the Nordic contest

The Phase 1 EUR 1.5 billion figure covers land acquisition, site preparation, four data hall buildings and the initial 110 MW substation. Pure Data Centres' construction schedule targets first-power in Q3 2027 and full commissioning of Phase 1 by mid-2028. The 550 MW expansion is contingent on customer commitments and additional grid-connection studies with Fingrid, which the company says are in progress. The Seinäjoki campus is Pure Data Centres' first purpose-built AI training and inference site. The Espoo and Stavanger campuses target enterprise cloud and hyperscale colocation workloads, while Seinäjoki is optimized for high-density GPU deployment, with mechanical cooling designed for direct-to-chip liquid cooling at the rack level. Rack power density targets are 130 kW per rack, roughly double the 60 kW density that the two older Pure Data Centres sites support.

"Galaxy Bets Its Data Center Edge on Power, Not 'Bragawatts'." — Data Center Knowledge, Aug 14

The build follows Meta's decision in May to consolidate its Nordic AI training footprint at the Odense, Denmark campus rather than expanding in Finland, and comes as Microsoft, Google and Amazon compete for the remaining renewable-anchored Nordic sites. Pure Data Centres is a portfolio company of Oaktree Capital Management, which has funded roughly $3 billion of data-center commitments since 2023. Finland's regulatory environment has accelerated data-center approvals since the Ministry of Economic Affairs and Employment adopted the streamlined permitting framework in January 2026. Under the new framework, projects above 100 MW receive expedited environmental review and grid-connection studies, cutting typical permitting timelines from 24 months to nine. Pure Data Centres' Seinäjoki application was filed in April and received conditional approval in early August, consistent with the streamlined timeline.

The Data Center Knowledge round-up also flagged Galaxy Digital's Project Merlin in McGregor, Texas, a 74 MW Phase 1 build on 500 acres of newly acquired land. Galaxy Digital's approach centers on grid-model optimization rather than nameplate power, with CEO Mike Novogratz telling the trade publication his team is betting on power, not "bragawatts." The Merlin build reflects the same pattern of hyperscale-adjacent operators moving from raw capacity to grid-integration expertise as the primary competitive moat.

ApplicationEconomic

Uber and Pony.ai commit 2,000 robotaxis to five European cities in first Chinese-AV EU scale-out

TechCrunch · 2026-08-14 · Source ↗

Uber and Nasdaq-listed Pony.ai expanded their 2025 partnership from Zagreb to four additional unnamed European cities, committing to more than 2,000 Level 4 robotaxis under a phased plan that also extends into the Middle East. Neither cities nor timing disclosed.

Uber and Nasdaq-listed Pony.ai on Friday announced a scaled expansion of the robotaxi partnership the two companies signed in May 2025, committing to deploy more than 2,000 vehicles across five European cities under a phased plan that also extends into unnamed Middle East markets. The initial commercial service, launched earlier this year in Zagreb, Croatia, will expand to four additional European cities that the two companies declined to name. Neither company disclosed a deployment timeline or capital commitments, saying details will emerge in phases.

The joint deployment model, described in a Business Wire release timestamped 06:12 UTC, combines Pony.ai's Level 4 autonomous stack with the Uber app front end and delegates fleet maintenance, cleaning and charging to local operating partners on a market-by-market basis. Ownership of the underlying rolling stock varies. In Zagreb, Uber operates as an asset-light platform and a local mobility company holds title to the fleet. In two of the four unnamed European markets, Pony.ai owns the vehicles directly and leases them to Uber-adjacent operating partners.

"Details will be revealed in phases." — Joint Uber-Pony.ai release

The first Chinese-AV EU deployment at meaningful scale

The commitment lands three days ahead of Pony.ai's Q2 earnings release on Monday, August 18. The company reported RMB 236 million ($34.7 million) in Q1 revenue at 145 percent year-on-year growth, with robotaxi revenue at RMB 59.12 million ($8.7 million) up 395.4 percent year-on-year. Management raised the 2026 fleet target from 3,000 to 3,500-plus vehicles on the Q1 call, and Friday's Europe announcement effectively adds another 2,000 vehicles on top of that number, though the delivery cadence remains open.

"The most consequential deployment of Chinese autonomous technology outside China." — Reuters analysis

The Europe commitment marks the first time a Chinese autonomous-vehicle developer has deployed at meaningful scale inside the European Union. Pony.ai holds no California Public Utilities Commission driverless deployment permit, has zero operating vehicles in the United States and remains under the Biden-era executive order that restricts Chinese connected-vehicle software from the US market. Europe has no equivalent restriction, and the EU AI Act's high-risk provisions that took effect August 2 apply to autonomous driving under Annex III but do not prohibit foreign developers.

Uber's position leans on its European licensing surface. The company operates in 27 of the 27 EU member states plus the UK, Norway and Switzerland, and holds an active platform license under the Digital Services Act since February 2024. Uber will manage rider onboarding, ride matching, pricing, safety escalation and insurance policies through its existing app. Pony.ai supplies vehicle software, remote monitoring from a Guangzhou command center staffed under EU-approved third-country data-transfer language, and remote-assistance operators. Bernstein estimated vehicle capex of $180,000 per robotaxi including the fifth-generation sensor kit, versus Waymo's Zeekr-based platform at approximately $135,000 per unit. Pony.ai's cost-of-goods advantage historically came from lower Chinese-market vehicle sourcing and lower labor costs in remote assistance. Europe erases both. The German Bundeskartellamt has an active preliminary inquiry into the Pony.ai-Uber Zagreb service under German merger control, though no formal proceeding is docketed.

ApplicationPolicy

Apple trains proprietary China AI model with Alibaba, first foreign firm to clear CAC approval

MacRumors / Reuters · 2026-08-14 · Source ↗

Apple has trained a proprietary large language model for the Chinese market with training help from Alibaba's Qwen team, three sources tell Reuters. First foreign-firm approval from Beijing under CAC algorithm-filing regime. Rollout targets iOS, iPadOS, macOS and visionOS across mainland China in the coming months.

Apple has trained a proprietary large language model dedicated to the Chinese market with training assistance from Alibaba's Qwen team, three people briefed on the arrangement told Reuters in a story that broke Friday morning Beijing time. The move ends nearly two years of Apple attempting to route Apple Intelligence in China through pure third-party inference and positions Apple as the first foreign company to receive Cyberspace Administration of China approval for a proprietary generative-AI service inside the country.

The China model runs on-device where possible and offloads to Alibaba Cloud for larger inference requests. The Chinese CAC granted approval for the on-device tier through its algorithm-filing regime, one of only two foreign companies to secure approval alongside Samsung, though Samsung's approval covers a locally routed Gemini instance and not a first-party model. Apple's approval marks the first time a foreign developer has registered its own model in the CAC catalog under the 2023 generative-AI interim measures.

"Bringing Apple Intelligence to our largest overseas market is a top priority for the company through the balance of calendar 2026." — Tim Cook, Apple July 31 earnings call

Two years of delay end in a first-party build

The China rollout was originally scheduled for early 2025 and slipped repeatedly. Apple had planned to route Apple Intelligence traffic in the mainland market through Baidu's ERNIE and then through Alibaba's Qwen. Both arrangements ran into CAC review friction over training-data provenance and content-filter compliance. Executives inside Cupertino concluded in Q3 2025 that only a first-party model with Chinese-labeled training data would clear approval, and Apple assigned an approximately 240-person team split between Beijing, Shanghai and Cupertino to train and align the model with Alibaba's on-site help.

"Apple is the first foreign company approved to offer a proprietary AI model in the country." — Reuters

The Alibaba relationship is technical rather than commercial. Alibaba's Qwen team provided training infrastructure, safety-alignment scaffolding and a Chinese-language pretraining corpus. Apple contributes the base architecture, tokenizer and fine-tuning methodology derived from its Apple Foundation Models series. Alibaba does not receive licensing fees per query, and Apple owns the resulting model weights outright. Alibaba Cloud does host the offload tier under a services agreement whose terms remain undisclosed. Rollout of the China-only Apple Intelligence experience is expected in the coming months, in time for the iPhone 18 launch cycle. Features include on-device summarization, notification triage, writing tools and image generation constrained by CAC content categories. Cloud offload will route to Alibaba Cloud data centers inside mainland China, and no user data will leave the mainland, according to the disclosure filed by Apple's China subsidiary. That structure differs from the Private Cloud Compute architecture Apple uses in every other market.

The regulatory precedent matters more than the product. Apple executives briefed the White House this week on the China arrangement, and the disclosure caused visible discomfort in the Trump administration's ongoing China-tech strategy discussions. The Bureau of Industry and Security is separately reviewing whether the model's training on H20-tier Nvidia hardware inside China qualifies as an export-controlled activity under the January 2025 rule updates. Apple's Chinese smartphone market share dropped to 14.1 percent in Q2 2026 from 17.2 percent a year earlier, per IDC. Huawei retook the number-one position at 18.9 percent, with Mate 80 Pro featuring an on-device Pangu model trained on Huawei's Ascend 910C hardware. Xiaomi at 15.4 percent shipped its own MiMo-Turbo. Alibaba shares gained 4.7 percent in Hong Kong trading Friday on the news, valuing the company at $412 billion. Baidu, whose ERNIE was originally selected for Apple Intelligence and then quietly dropped, closed down 2.3 percent in Hong Kong.

SafetyPolicy

France's DGFiP confirms tax-authority breach exposing up to 2 million taxpayer records via MFA bypass

The Register · 2026-08-14 · Source ↗

France's Directorate General of Public Finances confirmed Friday that a threat actor using the alias ZeroBytes accessed its systems in late June through stolen credentials plus a novel MFA bypass technique, extracting personal and financial data for up to 2 million French taxpayers. Public disclosure came 47 days after intrusion and only after ZeroBytes advertised the data on a criminal forum.

France's Directorate General of Public Finances confirmed Friday that a threat actor operating under the alias ZeroBytes accessed the agency's systems in late June through a combination of stolen credentials and a multi-factor-authentication bypass technique, extracting personal and financial data for up to 2 million French taxpayers. The disclosure came 47 days after the initial intrusion and only after ZeroBytes advertised the stolen dataset on a well-known Russian-language cybercrime forum on August 12.

The breach is one of the largest single-day disclosures of French public-sector data on record and lands in the middle of an active French government push to embed AI into tax-collection workflows. DGFiP has been rolling out an AI-assisted revenue-verification system called ARBOR since 2024 that cross-references tax filings against real-estate registers, employer withholding data and financial-account information. The exposed records include the exact inputs and outputs of that verification pipeline for a substantial fraction of French taxpayers.

"Access had been severed at the end of June as part of an audit." — DGFiP official statement

Novel MFA bypass and 47-day disclosure gap

Data exposed in the breach includes internal tax identifiers, full civil names, dates and places of birth, postal addresses, family compositions, number of tax dependents, reference tax income (revenu fiscal de reference) and withholding tax rates. The Register reported the breach affected approximately 2 million French taxpayers, both individuals and professionals, though FrenchBreaches placed the figure closer to 700,000 records at 678,438 confirmed. The discrepancy reflects an ongoing DGFiP internal audit; ZeroBytes claimed 2 million records on the criminal forum, and DGFiP has not yet confirmed the total.

"A novel MFA bypass technique." — DGFiP description of the attack chain

The MFA bypass technique used against DGFiP has drawn particular attention. The attacker obtained legitimate credentials through a credential-stuffing or targeted-phishing operation and then chained those credentials with what DGFiP described as "a novel MFA bypass technique" that let the threat actor complete the second-factor challenge without possessing the underlying MFA token. The technique is under investigation by ANSSI, France's national cybersecurity agency, and Interior Ministry cybercrime prosecutors. Investigators have not publicly disclosed whether the bypass exploited a session-token replay vulnerability or a browser-in-the-middle attack against DGFiP's authentication flow.

The 47-day disclosure gap has become a political issue in Paris. DGFiP detected the unauthorized access during routine security audits in late June, severed the access and made no public disclosure until Friday, when ZeroBytes' criminal-forum posting forced the agency's hand. Finance Minister Amelie de Montchalin ordered a full internal review Thursday evening, and Prime Minister Elisabeth Borne's cabinet is scheduled to discuss the breach disclosure timeline at Monday's Council of Ministers meeting. The delayed disclosure appears to violate CNIL's 72-hour notification requirement under GDPR, and CNIL confirmed Friday it has opened a formal inquiry. The AI-adjacent dimension is the exposure of ARBOR training data: DGFiP has trained approximately 340 models over the past 18 months for tax-fraud detection and taxpayer-behavior scoring, drawing heavily from the exposed records. Cumulative disclosed victim counts across French public-sector breaches in 2026 exceed 22 million, and Friday's disclosure adds a further 2 million if the ZeroBytes claim proves accurate.