Daily 2026-07-16
TechApplication

Cadence ships AuraStack, the first agentic AI platform for PCB and advanced chip packaging

Cadence Design Systems opened AuraStack for orders on Wednesday, framing it as the industry's first agentic AI platform for printed circuit boards and advanced chip packaging. The platform runs on Nvidia hardware, taps TSMC as the co-development partner for advanced packaging convergence, and lets customers plug in Claude, ChatGPT, Gemini or open-source weights as the underlying model. Cadence claims up to a 50 percent cut in time-to-market and up to a 15-fold productivity lift on individual tasks, with the rollout completing in September.

Forbes · 2026-07-15 · Source ↗

Cadence Design Systems introduced AuraStack AI Super Agent on Wednesday, calling it the world's first agentic AI platform for printed circuit board and advanced chip packaging design. Engineers describe goals in plain English. AuraStack plans a workflow, calls the existing Cadence Allegro X and Integrity 3D-IC tools underneath, lays out the routing, virtually tests the design, and hands back a completed file. Cadence CEO Anirudh Devgan pitched it Wednesday as the delegation model coming to hardware: the human sets intent, the agent runs the physical layout task with the same tools an EDA engineer uses today.

The productivity claims are concrete. Cadence says AuraStack cuts overall time-to-market by up to 50 percent and drives up to 15-fold productivity gains on individual tasks such as trace routing, placement, and design-rule verification. Availability lands in stages this year, with the full rollout completing in September. Pricing was not disclosed. The platform runs on Nvidia GPUs for the compute layer, with Nvidia GB200 and CUDA-X libraries pre-integrated, and Cadence and TSMC signed a co-development partnership to accelerate advanced-package implementation through AI-driven automation, framed as helping customers hit "timely design convergence for increasingly complex multi-die systems."

"AuraStack is the world's first agentic AI platform for PCB and advanced packaging. It allows engineers to focus on high-value creative work while the agent handles the mechanical convergence steps."— Anirudh Devgan, Cadence CEO

Model-agnostic by design

The most striking product decision is that AuraStack does not lock customers to one model. Cadence customers can pair the agent with OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini or open-source models, then swap between them per task. Devgan told analysts the design intent, the tool integrations, and the physical simulation harness are all Cadence's; the language model is the interchangeable head. For enterprise EDA customers who have compliance reasons to run one model in one workflow and another elsewhere, that flexibility matters more than raw benchmark scores.

"Customers can pair AuraStack with the AI model of their choice — ChatGPT, Claude, Gemini, or open-source weights — because the physical design layer is what we own."— Cadence press release, July 15

Early customers and the Nvidia stack

Cadence confirmed Nvidia, TSMC, and Schneider Electric as early adopters. Nvidia's role is both customer and technology partner: AuraStack runs on Nvidia GB200 systems and taps CUDA-X libraries for the routing and simulation kernels. Schneider Electric is deploying AuraStack for its industrial-controls board designs. TSMC is using it inside its advanced-packaging convergence flows for HBM4 and CoWoS-L 2.5D stacks. The Register described the announcement as "AuraStack melds AI with HPC to speed PCB and advanced packaging design," pointing to the tight coupling between Cadence's simulator, Nvidia's compute, and the LLM planning layer.

Why this lands now

Every major chip designer is hunting for productivity gains as advanced-node designs get harder. TSMC's advanced packaging capacity is sold out through 2027. Nvidia holds most of it. Design cycles for HBM4 chiplet stacks and CoWoS-L are the bottleneck at the front end of the pipeline, and Cadence's argument is that AI agents can compress those cycles by half. The bear case is straightforward: virtual engineers still fail on the hard physical corners, and PCB design is one of those disciplines where a single missed impedance target means a respin. Cadence's counter is that AuraStack uses the same simulator and design-rule engines EDA engineers already trust; the agent orchestrates, it does not replace the ground-truth tool.

The Big Picture

Cadence opened orders on AuraStack Wednesday, calling it the first agentic AI platform for PCB and advanced chip packaging, running on Nvidia GB200 hardware with TSMC as the advanced-packaging co-development partner and Nvidia, TSMC and Schneider Electric confirmed as early customers. Cadence claims a 50 percent cut in time-to-market and 15-fold task-level productivity, and lets buyers pair the agent with Claude, ChatGPT, Gemini or open-source weights. The rollout completes in September.

Nvidia shipped the first H200 units into China Wednesday, US Under Secretary for Industry and Security Alan Estevez confirming the deliveries in testimony before the House Foreign Affairs Committee three years after the Biden administration first blocked them. Alibaba, Tencent and ByteDance are among the ten Chinese firms approved to buy under the January 13 rule. Microsoft's July Patch Tuesday landed 570 CVEs, the largest single Patch Tuesday on record, and Microsoft credited an internal AI-powered vulnerability discovery system for the surge, alongside three zero-days: an Active Directory Federation Services privilege bug already under active exploitation, a SharePoint Server RCE also being exploited, and a publicly disclosed BitLocker bypass. Nvidia unveiled the Blackwell Jetson Thor T3000 and T2000 modules at 865 and 400 FP4 teraflops with the Cosmos 3 Edge 4-billion-parameter robot foundation model landing alongside.

Neko Health closed a $700 million Series C at close to $7 billion led by Lightspeed with Zuckerberg, Priscilla Chan, will.i.am, Zoë Saldaña and Alexis Ohanian on the cap table, funding Daniel Ek's Manhattan clinic launch and a four-fold valuation jump from January 2025. Indian AI coding startup Emergent became a unicorn at $1.5 billion on a $130 million Series C led by Creaegis, at $120 million ARR sixteen months after launch. Amsterdam bricklaying-robot firm Monumental raised $32 million Series B led by Khosla to enter the US. Anthropic tagged Claude Code v2.1.210 with a live elapsed-time counter on tool calls and hardened the Agent tool against indirect prompt injection, made HIPAA self-serve for Enterprise and API tiers, and shipped Microsoft 365 connector write tools that let Claude draft and send email, manage calendars, and update OneDrive and SharePoint files. ggml-org shipped ten llama.cpp builds through the UTC day covering SYCL Flash Attention on Intel Battlemage with 1.21x prefill speedup, CUDA lightning-indexer kernels, DeepSeek V4 graph-split reductions, and an OpenCL Adreno A7x MoE-kernel workaround. Kyutai's MuScriptor open multi-instrument music transcription model topped Hugging Face daily papers at 556 upvotes with VLRLab-OCR's MonkeyOCRv2 at 144.

More today

20 stories · 2026-07-15
TechPolicy

Estevez confirms Nvidia H200 shipments into China have begun, three years after the initial block

Voice of Emirates · 2026-07-15 · Source ↗

US Under Secretary for Industry and Security Alan Estevez told the House Foreign Affairs Committee Wednesday that first H200 units have arrived in China under the January 13 rule, ending a three-year block. Alibaba, Tencent and ByteDance sit among the ten approved buyers.

US Under Secretary of Commerce for Industry and Security Alan Estevez confirmed Wednesday in testimony before the House Foreign Affairs Committee that first shipments of Nvidia H200 accelerators have arrived in China. The confirmation ends a three-year policy stand-off dating to the October 2022 Biden-era controls and their October 2023 tightening, which had blocked the H200's export path into the country entirely. The reversal traces to President Trump's December 2025 announcement that Nvidia would be permitted to sell H200 chips to China, codified into a formal Bureau of Industry and Security rule on January 13, 2026.

The commercial stakes are large. The H200 is roughly six times more powerful than the H20, the compliance chip Nvidia designed for the previous export-control regime. Tom's Hardware reported earlier this month that Nvidia is preparing shipments of up to 82,000 GPUs, subject to a 25 percent tax under the licensing arrangement. The Department of Commerce approved requests in May from about ten major Chinese firms, including Alibaba, Tencent and ByteDance, to acquire H200 accelerators. Actual deliveries had not commenced by July 1. Estevez's Wednesday statement is the first official confirmation from a sitting Commerce official that shipments have moved.

"Shipments of Nvidia H200 chips have already begun arriving in China."— Alan Estevez, US Under Secretary for Industry and Security

What still stands and what the reversal buys

The advanced-node chips remain restricted. B100, B200, GB200 and Blackwell Ultra products cannot be shipped to China under current rules. The H200 sits one notch below those in Nvidia's product ladder but comfortably above the H20 the H200 replaces on the export list. Analysts frame the reversal as a bet that keeping Nvidia's ecosystem, CUDA and NVLink dominant inside Chinese data centers is worth ceding some raw performance ground, versus the alternative in which Chinese hyperscalers accelerate their pivot to Huawei Ascend and homegrown accelerators. The Brookings piece last month captured the counterargument bluntly: with restrictions on the top of the stack still in place, the H200 flow slows but does not stop Chinese sovereignty pushes on inference silicon.

For Nvidia's revenue mix, the reopening is material. China accounted for roughly 25 percent of data-center revenue at peak. The 25 percent tax pockets some of the upside for the US Treasury, but the shift widens Nvidia's addressable market at a moment when the company's global backlog was already exceeding capacity. On Wednesday's headlines, Nvidia shares rose in New York trading. The company has not commented publicly on shipment volumes.

SafetyTech

Microsoft ships record 570-CVE July Patch Tuesday, credits AI-powered vulnerability discovery for the surge

gHacks · 2026-07-15 · Source ↗

The largest single Patch Tuesday drop on record fixes 59 critical bugs, three zero-days including an ADFS privilege bug and a SharePoint RCE already under active exploit, and a Copilot RCE (CVE-2026-48561). Microsoft attributes the flaw count to an internal AI-powered vulnerability discovery system running across the Windows codebase.

Microsoft shipped the July 2026 Patch Tuesday drop Tuesday, addressing 570 vulnerabilities across the Windows and Microsoft product stack. That is the largest count Microsoft has ever landed in a single monthly release. The prior high sat under 300. The breakdown lands 254 elevation-of-privilege bugs, 145 remote-code-execution bugs, 102 information-disclosure bugs, 35 denial-of-service, 17 security-feature-bypass, and 16 spoofing bugs. Fifty-nine of the 570 are rated Critical.

Three zero-days are on the list. CVE-2026-56155 in Active Directory Federation Services allows attackers to gain administrative privileges through insufficient access-control granularity and is already under active exploitation. CVE-2026-56164 in Microsoft SharePoint Server permits remote privilege escalation via missing authentication on a critical function, also being exploited in the wild. CVE-2026-50661 is a publicly-disclosed BitLocker bypass that permits an attacker with brief physical access to reach encrypted data. Windows 11 users get the fixes via KB5101650 and KB5099414; Windows 10 Extended Security Updates arrive via KB5099539.

"A new AI-powered vulnerability discovery system that identifies security flaws across the Windows codebase before attackers can exploit them."— Microsoft, on the record patch count

Copilot enters the CVE list

CVE-2026-48561 patches a critical-severity remote-code-execution bug in Microsoft Copilot. Full technical details are held back per Microsoft's coordinated disclosure practice. The item continues 2026's running theme of Copilot-branded features appearing in the Patch Tuesday releases as attackers probe the LLM-integration surface for prompt-injection paths that reach real code execution. gHacks noted the RCE marked the first time an AI assistant CVE has cleared the critical bar in a Patch Tuesday release since the classification was added last year.

Why the count spiked

Microsoft explicitly tied the surge to internal automation. The company said an AI-powered vulnerability discovery system now runs continuously against the Windows codebase, surfacing flaws faster than the manual security research pipeline previously could. Whether attackers have equivalent tooling is now the operational question. Independent researchers have noted for months that automated LLM-driven vulnerability discovery is inside the arsenal of both defenders and offensive actors; Microsoft's decision to disclose that it uses one at the source-tree level is the first public admission from a major vendor. The knock-on question is patching cadence. IT teams that manage large Windows fleets already struggled with the 200-plus CVE months. A record 570 in one drop lands on the same operators, and Microsoft has not announced pipeline changes to help enterprise deployment ride the new pace.

Tech

Nvidia unveils Blackwell Jetson Thor T3000 and T2000 with Cosmos 3 Edge robot foundation model

Nvidia · 2026-07-15 · Source ↗

The T3000 delivers 865 FP4 teraflops in half the size and power of the T5000 with 32 GB LPDDR5X and 273 GB/s bandwidth. Cosmos 3 Edge lands as a 4-billion-parameter foundation model for on-device robot vision, and new Jetson agent skills cut memory use by up to 15 GB across applications.

Nvidia announced two new Blackwell-based Jetson Thor modules Wednesday: the T3000 and the T2000. Both are pitched at moving robotics and edge-AI workloads from the developer-kit stage to mass-market deployment. The T3000 delivers 865 FP4 teraflops of AI compute in a form factor Nvidia describes as roughly half the size and half the power of the flagship Jetson T5000. Its internal composition pairs a Blackwell GPU with an eight-core Neoverse Arm CPU, 32 GB of LPDDR5X, 273 GB/s of memory bandwidth, and 25 GbE connectivity. The T2000 lands at 400 FP4 teraflops with 16 GB of memory for broader edge-AI applications where visual agents or lower-throughput autonomous systems are the target.

The software side is the interesting layer. Nvidia introduced Cosmos 3 Edge, a 4-billion-parameter robot foundation model that provides real-time vision analysis and on-device reasoning for embodied systems. The company also announced Jetson agent skills that automate stack-level optimization and cut memory usage by up to 15 GB across applications, enabling deployment on lower-memory configurations without performance loss. That is a meaningful lever: for humanoid and autonomous mobile robot customers, memory footprint per model is often the ceiling on how many concurrent models the platform can host.

"Jetson Thor is the ultimate platform for physical AI. The new T3000 and T2000 modules bring Blackwell compute into a form factor that lets partners ship at scale."— Nvidia, official product blog

Availability, price, and partner list

T3000 emulation lands this month via JetPack 7.2.1; T2000 support follows later this quarter. Physical modules become available in Q1 2027. Nvidia listed the existing Jetson AGX Thor developer kit at $3,499. The company published a partner ecosystem numbering more than 15 hardware companies including ADLINK, Advantech, Connect Tech and Seeed Studio, plus software partners Antmicro, RidgeRun, and Neurealm. On the customer side, Nvidia named 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, and Techman Robot as adopters of the AGX platform for humanoid and industrial robotics deployments. Jetson AGX Thor is currently powering next-generation humanoid systems inside 1X's neo humanoid and Boston Dynamics' Atlas commercial line.

The context

The physical-AI compute market is Nvidia's stated growth axis for 2027, and Jetson Thor is the enabling silicon. The T3000 hits at a price and power point where mid-market industrial customers can specify Blackwell compute in a robot without pricing themselves out. Whether the T2000 finds a customer base against Qualcomm's Snapdragon Ride and the growing catalog of RISC-V edge accelerators is the open competitive question. On raw teraflops, T2000 leads. On developer-ecosystem depth, so does the entire Jetson line.

Tech

Anthropic tags Claude Code v2.1.210 with a live elapsed-time counter, worktree isolation fix, and Agent-tool prompt-injection hardening

Releasebot · 2026-07-15 · Source ↗

Long-running tool calls now visibly tick in the collapsed summary line so they no longer look stuck. A worktree-isolation bug where subagents ran git commands against the main repo is fixed. The Agent tool has new defenses against indirect prompt injection.

Anthropic tagged Claude Code v2.1.210 Wednesday. The user-facing feature is a live elapsed-time counter attached to the collapsed tool-call summary line, giving long-running tool calls a visible tick so operators no longer have to guess whether the agent is working or wedged. The change addresses one of the top-voted complaints in the Claude Code GitHub tracker over the past month. For agent operators running multi-hour builds or benchmark suites through the CLI, the counter also functions as a rough progress signal.

The bigger changes sit beneath the UI. Anthropic fixed a class of bugs where isolation: 'worktree' subagents were executing git commands against the main repository rather than their isolated worktrees. That was a real correctness issue for the growing set of teams using the worktree pattern to run multiple agents on the same codebase in parallel: the isolation guarantee was silently broken for git operations, meaning a git-commit call inside subagent A could land against the main worktree. The v2.1.210 fix returns the isolation to its documented behavior.

"Hardened the Agent tool against indirect prompt injection vulnerabilities."— Claude Code v2.1.210 release notes, July 15

Indirect prompt injection

The security highlight is the release note "Hardened the Agent tool against indirect prompt injection vulnerabilities." Anthropic did not publish exploit details. The class of bug involves an attacker inserting malicious instructions into content the agent reads (a document, a webpage, a file the agent Grep's) that the LLM then treats as commands. Claude Code's Agent tool sits at a high-privilege position in the tool ecosystem: it can spawn subagents, call other tools, and write to disk. Hardening the surface against injection paths is baseline hygiene, and Anthropic's willingness to ship the fix in a point release rather than gate it to a major indicates the team is treating agent-tool injection as an active area rather than a latent risk.

The full list

Other fixes in v2.1.210: resolved "job not found" and "agent is still starting" errors during session transitions; fixed background workers, MCP servers, and approval mechanisms for stronger stability; corrected hook callback timeouts previously being misreported as user rejections; screen-reader mode now announces permission mode changes audibly; Agent view improvements with persistent selection marking; Grep pagination bug fixed; Windows plugin cache failures resolved; background-worker crash-loop scenarios patched; late-appearing symlinks now reconciled into sandbox deny-write lists. Anthropic also added startup warnings for deprecated permission-rule syntax, nudging users toward Edit(path) and Read(path) forms. The upgrade path is a standard npm bump; no manifest changes required for existing agent scripts.

ApplicationTech

Anthropic ships self-serve HIPAA, Microsoft 365 write tools, and Government Community Cloud access in one drop

Anthropic release notes · 2026-07-15 · Source ↗

Enterprise and API admins can now review the BAA and enable HIPAA compliance in one flow. Claude's Microsoft 365 connector gains write capability across Outlook, Calendar, OneDrive and SharePoint. GCC customers get access to Claude via the Microsoft 365 admin center.

Anthropic pushed a compliance-and-integration drop on Claude Wednesday covering three enterprise surfaces. HIPAA configuration is now self-serve. Eligible administrators on both the Claude Enterprise product and the Claude Platform (API) can review the Business Associate Agreement, download the implementation guide, and enable HIPAA in a single flow. Previously the process required manual paperwork with the Anthropic legal team, which added meaningful lead time for hospitals, insurers, and clinical software vendors that wanted Claude access with a signed BAA in place.

The Microsoft 365 connector gained write tools. Claude can now draft, send, and organize email, manage calendar events, update mailbox settings, and create and update files in OneDrive and SharePoint. Read and search work as before, and Teams remains read-only. Anthropic requires a Microsoft Entra administrator to consent to the updated permission set before organization members can flip write access on. This is the connector's first move past read-only mode and directly targets the daily-workflow surface where Copilot has been the default; Claude now lives inside the same Outlook-Calendar-OneDrive layer.

"Claude adds write tools for the Microsoft 365 connector to draft email, manage calendar events, and update OneDrive and SharePoint files."— Anthropic release notes, July 15

Government Community Cloud

A new Microsoft 365 admin-center setting is rolling out that allows non-federal customers in Government Community Cloud (GCC) to use Anthropic models directly. Public-sector customers who run inside GCC do not need federal-tier authorizations to run Claude against internal documents and mail, opening state and municipal buyers as an addressable segment through the same Microsoft admin surface enterprise customers already administer. The rollout begins Wednesday and is scheduled to complete over the next two weeks.

Focus mode and monthly recap

Anthropic also flipped on monthly recap and focus settings for Claude web and Desktop, first previewed on July 9. Monthly recap surfaces a personal Wrapped-style usage view. Focus mode gives users optional break reminders and quiet hours drawn from the 4D AI Fluency Framework Anthropic published earlier this quarter. The mode ships in beta for Free, Pro, and Max plans.

TechApplication

Google ships Nano Banana 2 Lite image model and lifts Gemini Omni Flash into API public preview

Google Gemini release notes · 2026-07-15 · Source ↗

Nano Banana 2 Lite launches as the fastest and most cost-efficient Gemini image model. Omni Flash lands in API public preview for native multimodal video generation. ATL Saathi, a Gemini-powered AI planning assistant, pilots in 100 Indian Atal Tinkering Lab schools.

Google shipped two developer-facing Gemini updates Wednesday, both aimed at cost and multimodal breadth. Nano Banana 2 Lite launches as Google's fastest and most cost-efficient image-generation model, positioned below Nano Banana 2 for use cases where per-image latency and per-token cost matter more than image fidelity. Google also brought Gemini Omni Flash to the API in public preview, describing it as a natively multimodal model that lets enterprise developers build custom, dynamic video workflows. The Omni Flash line pairs the vision-language backbone with an audio decoder capable of speech synthesis for accompanying narration.

The second wave is more narrowly scoped enterprise plumbing. Google Workspace users on the AI Expanded Access license tier now have higher usage limits on Fill with Gemini and the AI function inside Sheets, effective July 15. Workspace customers get promotional access through August 1 to higher limits on multi-slide presentation generation using Gemini in Google Slides. These are quota expansions rather than new capabilities, but they matter for teams that ran into throttles on the previous limits when generating tabular fills at scale.

"Nano Banana 2 Lite is our fastest and most cost-efficient Gemini image model. Omni Flash brings native multimodal video generation to enterprises and developers."— Google, Gemini release notes, July 15

ATL Saathi lands in Indian schools

The most visible outreach announcement is ATL Saathi, a Gemini-powered web app piloting in 100 Indian schools starting Wednesday. ATL Saathi acts as a 24/7 planning and training assistant for Atal Tinkering Lab educators. Atal Tinkering Labs is a NITI Aayog program that has established over 10,000 school-based innovation workshops across India since 2016. ATL Saathi gives lab facilitators a Gemini interface for curriculum planning, project idea generation, and troubleshooting on lab equipment. Google framed it as part of a broader push to embed Gemini inside Indian education infrastructure ahead of the July 17 launch of Gemini 3.5 Pro.

Independent of the Wednesday drop, Google is running a two-day countdown to Gemini 3.5 Pro. Leaked specifications include a 2-million-token context window, Deep Think reasoning gated behind the $250 Ultra tier, and API pricing landing near $1.25 input and $10 output per million tokens.

Application

OpenAI lifts ChatGPT custom instructions from 1,500 to 5,000 characters for paid tiers

Crypto Briefing · 2026-07-15 · Source ↗

Pro, Enterprise, Business and Education subscribers now have more than triple the previous character allowance for the persistent-instructions payload that shapes every response. Free users are excluded.

OpenAI raised the ChatGPT custom-instructions character limit from 1,500 to 5,000 on Wednesday for users on the Pro, Enterprise, Business, and Education tiers. The Free tier remains at 1,500. The change applies retroactively to existing custom-instruction payloads: previously-saved instructions continue to work, and users can now edit them out to the new ceiling. The custom-instructions feature became universal across chats in November 2025, meaning the block is applied to every conversation rather than requiring per-chat re-entry. The Wednesday change makes the payload substantially more expressive.

The move aligns with OpenAI's transition to GPT-5.1 across ChatGPT. OpenAI reports the newer model handles long custom-instruction payloads more reliably than its predecessors, retaining the specified voice, persona, and formatting constraints across long conversations without the drift that showed up on GPT-5.0. Crypto Briefing characterized the previous 1,500-character allowance as "roughly the length of a detailed text message" and the new 5,000 as "a proper briefing document."

"Custom instructions now let you save up to 5,000 characters, up from 1,500."— OpenAI, ChatGPT release notes

Why the size matters

For power users, the immediate consequence is space to write in-role directives that were previously abbreviated for the 1,500-character cap. Enterprise deployments use custom instructions to encode role definitions, policy constraints, and formatting requirements. A 5,000-character allowance lets a Legal team write a full compliance disclaimer regime; a Support team can encode a full tone-of-voice and knowledge-boundary policy; a Research team can specify the exact citation style. The change quietly shifts more of the practical prompt-engineering surface into a persistent per-user payload that never has to be reasserted per conversation.

Tech

ggml-org lands ten llama.cpp builds through the UTC day, headlined by SYCL Flash Attention with a 1.21× prefill speedup

llama.cpp releases · 2026-07-15 · Source ↗

b10016 ships SYCL Flash Attention with XMX engine via oneDNN. b10032 adds a CUDA lightning-indexer wmma kernel. b10021 is a DeepSeek V4 graph-split reduction. b10034 excludes MoE kernels on Adreno A7x to work around compiler miscompilation.

ggml-org tagged ten llama.cpp builds through the UTC day on July 15, landing between 09:10 UTC (b10016) and 20:34 UTC (b10034). The headline is b10016: SYCL Flash Attention with the XMX engine via oneDNN, delivering a claimed 1.21× prefill speedup at certain sequence lengths on compatible Intel Battlemage hardware. The XMX engine is Intel's matrix-multiplication acceleration path on Arc B-series and Battlemage discrete GPUs. Wiring Flash Attention to it via oneDNN is the SYCL backend's equivalent of the CUDA cuBLAS Flash Attention integration on Nvidia parts.

The rest of the day fills the surrounding SYCL and CUDA gaps. b10017 raised the minimum USM system-allocation buffer threshold from 1 GiB to 4 GiB to enable more aggressive VRAM overcommit. b10018 added a SYCL xielu activation-function operator. b10019 implemented SYCL fp16 kernel type support for conv2d_dw operations. b10020 fixed SYCL get_rows for Q2_K, Q4_K and Q5_K quantization formats. b10025 relaxed CUDA tensor-contiguity requirements for quantized concatenation operations and added test cases for non-contiguous scenarios.

"opencl: exclude some moe kernels on Adreno a7x — addresses compiler miscompilation issues affecting MoE model weight handling on certain A7x devices."— llama.cpp b10034 release note

DeepSeek V4 and lightning-indexer

b10021 lands a DeepSeek V4 graph-split reduction optimization. DeepSeek V4 shipped with a 1.6-trillion-parameter Pro variant and 284-billion-parameter Flash variant in late June with a fine-grained sparse-attention architecture that stresses the KV-cache and graph-split heuristics in llama.cpp's CUDA backend. b10021's reduction targets the memory-transfer overhead in that architecture's forward pass. b10032 adds a CUDA implementation of the GGML_OP_LIGHTNING_INDEXER operation, providing both generic vector and wmma kernel variants and relaxing the architecture requirement to Turing compatibility. Lightning indexer is the primitive DeepSeek V4 uses for sparse attention lookups; landing it on CUDA broadens the hardware base for full V4 inference.

Housekeeping

b10031 dropped the --stdin mutual-exclusion check from the tokenize tool to align CLI behavior with completion scripts. b10034 excluded some MoE kernels on Adreno A7x mobile GPUs to work around a Qualcomm compiler miscompilation issue that produced silently wrong outputs on MoE model weight handling. The Adreno path is llama.cpp's route into Android inference on Snapdragon 8 Elite silicon, so the workaround protects a growing user base until Qualcomm's compiler fix lands. Georgi Gerganov has flagged July 21 as the target for the Hexagon NPU code drop that closes out the Snapdragon inference story.

Tech

Kyutai's MuScriptor tops Hugging Face daily papers at 556 upvotes with open multi-instrument music transcription

Hugging Face papers · 2026-07-15 · Source ↗

Kyutai released MuScriptor Wednesday: an open-weights model that transcribes polyphonic multi-instrument audio into standard music notation, positioned as the open-source counterpart to closed proprietary systems.

Kyutai's MuScriptor topped Hugging Face daily papers Wednesday at 556 upvotes, more than triple the runner-up. The paper introduces the first open-weights model that transcribes polyphonic multi-instrument audio into standard Western music notation end-to-end. Prior open work on automatic music transcription had either constrained inputs to solo piano or produced piano-roll representations that require downstream conversion; MuScriptor produces score-ready notation directly, handling ensemble arrangements including strings, brass, woodwinds, and percussion in the same forward pass.

The Paris-based lab framed the release as continuous with its earlier open speech and audio work: an insistence on open weights, an insistence on model sizes small enough for practical inference on consumer hardware, and a permissive license. The upvote count is unusual. Hugging Face daily papers typically caps around the 100–200 range for top posts. MuScriptor's 556 puts it in the top handful of open releases this quarter by community reception. Musicologists, transcription workflow builders, and open-source music-production communities have all weighed in on the launch thread with early notation samples that community members characterize as production-usable on the ensemble tracks they tested.

"MuScriptor is the first open, general-purpose multi-instrument music transcription model. We release weights, code, and a live demo."— Kyutai, MuScriptor release

Position in the transcription stack

Closed offerings from Klangio, Anthropic Music Reader, and Google MagentaScribe have dominated the multi-instrument space to date. MuScriptor's open release breaks that. Kyutai did not publish head-to-head benchmarks against the closed systems, but the paper does report state-of-the-art scores on the Bach10, MedleyDB, and MAESTRO benchmarks for the open category. Community members have begun benchmarking against Klangio Advanced on standard test tracks; early comparisons suggest MuScriptor sits within a few percentage points of Klangio on ensemble accuracy and above it on rare-instrument coverage.

Tech

VLRLab-OCR ships MonkeyOCRv2 as an open visual-text foundation model for document AI at 144 upvotes

Hugging Face papers · 2026-07-15 · Source ↗

MonkeyOCRv2 unifies text detection, layout parsing, table and formula recognition and multi-page reading order in one open-weights model, targeting the same use cases Nougat and PaddleOCR carve up.

Huazhong University of Science and Technology's VLRLab-OCR released MonkeyOCRv2 Wednesday, a visual-text foundation model for document AI that unifies text detection, layout parsing, table structure recognition, formula recognition, and multi-page reading-order inference in a single model. The v2 update sits atop the original MonkeyOCR released in early 2025 and roughly doubles document-understanding benchmark scores on DocVQA and OmniDocBench. Weights are permissively licensed on Hugging Face.

The unification story is what mattered to the community. Existing OSS document AI pipelines chain together separate models for each stage: PaddleOCR for text detection, LayoutLMv3 or PP-StructureV2 for structure, and Nougat or MathPix for equation extraction. Every hand-off is a place for errors to compound. MonkeyOCRv2 runs the entire pipeline through one visual-language transformer, and VLRLab-OCR reports substantial accuracy improvements on complex scientific papers precisely because the model treats formulas, tables, and body text as a single joint decoding problem rather than a sequence of independent tasks.

"MonkeyOCRv2 handles detection, layout, tables, and formulas in one pass. No pipeline glue."— VLRLab-OCR, MonkeyOCRv2 release

Where it lands

OCR-heavy verticals — legal document processing, academic paper ingestion, insurance-claim scans, healthcare records — have been waiting for an open-weights document-AI model that matches closed offerings from Amazon Textract, Google Document AI, and Anthropic Vision. MonkeyOCRv2 lands in that gap. The 144-upvote count on Hugging Face daily papers reflects heavy interest from RAG builders who now have a plug-in visual-text model for PDF ingestion pipelines. Anthropic's Retrievo team and LlamaIndex both put out early benchmarking threads on the launch day.

Tech

ServiceNow-AI's SynthDocBench introduces a controlled benchmark for long-context visual document understanding at 52 upvotes

Hugging Face papers · 2026-07-15 · Source ↗

ServiceNow-AI released SynthDocBench Wednesday: a benchmark generator that lets researchers control exactly how long, how visually complex, and how information-dense each test document is, isolating why visual language models fail on long documents.

ServiceNow-AI's SynthDocBench, published Wednesday and sitting at 52 upvotes on Hugging Face daily papers, targets a problem that has been quietly holding back document-AI evaluation: existing long-document benchmarks conflate too many variables. Longer documents in the wild are also more visually complex, denser with tables and figures, and covered by different distributions of information layouts. When a VLM's accuracy drops on longer documents, the researcher cannot say whether the drop came from length itself, visual complexity, information density, or the interaction between them.

SynthDocBench solves that by generating benchmarks synthetically with explicit control over each axis. The generator holds document length constant while varying visual complexity; it holds visual complexity constant while varying length; it holds both constant while varying information density. ServiceNow-AI then ran an ablation across a dozen frontier VLMs on the resulting benchmark. The findings are striking: visual complexity, not length, is the dominant failure axis. VLMs handle much longer documents than the community has been benchmarking on, provided visual complexity stays under a threshold.

"Longer documents don't fail because they're long. They fail because the visual complexity is compounding under the hood, and current benchmarks conflate the two."— SynthDocBench paper, abstract

What the benchmark unlocks

The synthetic-generation approach lets ServiceNow-AI publish a benchmark that will not saturate. Researchers can regenerate SynthDocBench with a fresh seed to produce a held-out test set immune to training-data leakage. That is a real problem in the current DocVQA/OmniDocBench era where high scores partially reflect training-corpus contamination. SynthDocBench's generator is open, so any lab can produce their own held-out variant while comparing to the paper's shared axes.

Tech

ByteDance Seed's "Read It Back" shows pretrained MLLMs work as zero-shot reward models for text-to-image generation

Hugging Face papers · 2026-07-15 · Source ↗

ByteDance Seed proposes reusing pretrained multimodal LLMs as text-to-image reward models with zero fine-tuning. The paper reports substantial improvement over trained reward models on standard image-alignment benchmarks at 42 upvotes.

ByteDance Seed's paper "Read It Back" argues that pretrained multimodal large language models already carry the alignment signal needed to serve as reward models for text-to-image generation, with zero task-specific training. The setup is elegant: given a candidate generated image and its text prompt, prompt a pretrained MLLM to describe the image. Compare the model's description back to the original prompt with an off-the-shelf text similarity metric. Use the score as the reward signal in reinforcement learning from AI feedback.

The empirical result is that this zero-shot reward matches or beats reward models specifically trained on human preference data over standard benchmarks including HPS v2, ImageReward, and Pick-a-Pic. That has two consequences for practitioners. First, teams building text-to-image RLAIF pipelines no longer need to invest in preference-labeling campaigns to bootstrap a reward model; they can plug in any strong MLLM. Second, the paper suggests that whatever visual-textual alignment the MLLM learned during pretraining transfers cleanly to the T2I preference-scoring task, which sharpens the argument that MLLMs learn richer alignment signals than benchmark scores capture.

"MLLMs already know how to evaluate text-to-image alignment. We just have to ask them to read the image back."— ByteDance Seed, Read It Back

Practical scope

The method has practical constraints. MLLMs used as rewards need to be strong enough that their descriptions are faithful; the paper's headline results use Qwen3.6-VL and GPT-4.1V. The reward-model latency also becomes non-trivial in RLAIF loops; each reward evaluation is an MLLM forward pass. The upside is the removal of a preference-data collection cycle. For labs training image generators in domains where preference-labeling is expensive — medical imagery, scientific visualization — this is the more meaningful benefit.

Tech

TIGER-Lab's "Search Beyond What Can Be Taught" evolves the knowledge boundary in agentic visual generation at 28 upvotes

Hugging Face papers · 2026-07-15 · Source ↗

TIGER-Lab published a method that lets agentic image generators actively search beyond their training distribution when a prompt asks for concepts the base model has not seen, producing higher-quality outputs on out-of-distribution prompts.

TIGER-Lab's Wednesday paper on evolving the knowledge boundary in agentic visual generation targets one of the most persistent failure modes of text-to-image systems: prompts that reference concepts, objects, or styles that the base model was not trained on. Standard practice is to accept the failure or fine-tune the model on domain data. TIGER-Lab argues for a third path: give the agent a search tool and let it retrieve reference imagery at generation time, then condition the diffusion process on the retrieved references. The paper describes the retrieval-conditioned generation pipeline and reports substantial quality improvements on out-of-distribution prompts spanning obscure architectural styles, rare biological species, and niche technical diagrams.

"An agentic image generator that hits an unfamiliar concept should look it up, the way a human illustrator would."— TIGER-Lab, Search Beyond What Can Be Taught

The paper sits alongside a second TIGER-Lab release on Wednesday, Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models, which won 12 upvotes. That paper argues coding agent foundation models benefit from a mid-training stage where the model learns function-aware fill-in-the-middle over real repository code, rather than jumping directly to instruction tuning. Combined, the two TIGER-Lab releases articulate a coherent research position: agentic behavior is best learned through targeted mid-training and inference-time retrieval, not through scaling alone.

Tech

Also on Hugging Face daily papers: EPFL Blind-Spots-Bench, Eternis probing LLM forecaster calibration, SJTU on repo knowledge acquisition

Hugging Face papers · 2026-07-15 · Source ↗

EPFL's Blind-Spots-Bench (22 upvotes) evaluates blind spots in multimodal models. Eternis publishes "What LLM Forecasters Know but Don't Say" (3 upvotes). Shanghai Jiao Tong University lands a QA-driven repository-knowledge acquisition paper for software issue resolution (6 upvotes).

EPFL's Blind-Spots-Bench, at 22 upvotes, is the day's clearest evaluation contribution. The team assembled 4,800 image-question pairs across 12 domains where multimodal models systematically fail: fine-grained spatial reasoning, specialized medical imagery, non-Western cultural artifacts, technical diagrams, and long-tail objects underrepresented in web-scale training. The paper reports that current frontier MLLMs including GPT-5.6-Sol, Claude Sonnet 5, and Gemini 3.5 Pro drop 15–30 percentage points below their reported general-benchmark scores on the blind-spot categories. The takeaway is that current benchmarks understate the gap between frontier VLMs and human performance because they under-sample failure-prone categories.

Eternis, a small alignment-focused lab, published "What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness" at 3 upvotes. The paper probes the internal representations of LLMs used as forecasters — models asked to predict future outcomes and give probability estimates — and finds that the models' internal activations often carry more calibrated uncertainty than the models express in their verbalized probability estimates. In other words, the models know they are uncertain but don't say so.

"LLM forecasters know when they're uncertain. They just don't tell you unless you look at the residual stream."— Eternis, on probing LLM forecaster calibration

Shanghai Jiao Tong and coding agents

Shanghai Jiao Tong University's "Know Before Fix: QA-Driven Repository Knowledge Acquisition for Software Issue Resolution" landed at 6 upvotes. The paper argues that coding-agent performance on SWE-bench-style benchmarks improves substantially when the agent first executes a structured question-answering pass over the target repository to build a knowledge graph, before attempting the fix. The team reports state-of-the-art results on SWE-bench Verified when Know Before Fix is applied on top of Claude Sonnet 5 as the underlying model. Combined with the TIGER-Lab function-aware fill-in-the-middle paper on the same day, the direction of travel in coding-agent research is clear: extract structure first, patch second.

EconomicApplication

Neko Health raises $700M Series C at nearly $7 billion for Manhattan clinic launch, four-fold up from January 2025

TechCrunch · 2026-07-15 · Source ↗

Daniel Ek's Neko Health closed a Lightspeed-led round with O.G. Venture Partners co-leading and Zuckerberg, Priscilla Chan, will.i.am, Zoë Saldaña and Alexis Ohanian among individual backers. The round funds a Manhattan preventive-scan flagship clinic opening this year.

Neko Health, the AI-driven body-scanning clinic company founded by Spotify co-founder Daniel Ek and Hjalmar Nilsonne, closed a $700 million Series C Wednesday at a valuation close to $7 billion. That is a roughly four-fold jump from the $1.7 billion post-money valuation Neko Health carried at its Series B in January 2025 and makes it one of the highest-valued Swedish tech companies in the country's history. Lightspeed Venture Partners led the round with O.G. Venture Partners co-leading. Existing investors Atomico, General Catalyst, and Lakestar joined, alongside new backers Liberty City Ventures, Positive Sum, and BDT & MSD.

The individual investor list is the celebrity cover for the round. Mark Zuckerberg and Priscilla Chan, Ari Emanuel, Claudia Schiffer, Danny Meyer, Jimmy Iovine, Maria Sharapova, Thierry Henry, Tim Ferriss, will.i.am, Alexis Ohanian from Seven Seven Six, Alex Tew and Michael Acton-Smith from Calm, Gary Vaynerchuk, Jessie Inchauspé, Katie Haun, Raj Shamani, Steven Bartlett, Zoë Saldaña and Marco Perego-Saldaña all wrote checks. The list functions as marketing surface for the US launch as much as capital.

"Neko will deploy the funding to fuel its immediate across-the-pond expansion, opening its inaugural flagship preventive scanning clinic in Manhattan later this year."— Neko Health, press statement

The scan

Neko's clinic model pairs a whole-body imaging scan — hundreds of sensors capturing skin, cardiovascular, and body-composition data — with same-day bloodwork. The AI pipeline processes the imagery and lab results into a preventive health report reviewed by an in-clinic physician. Clinics run in Stockholm and London today. The Manhattan flagship is Neko's first US location.

Preventive health as an AI application layer

Neko sits inside a broader wave of preventive-health startups using AI to make whole-body imaging economically viable. Competitors including Prenuvo, Ezra, and SimonMed have raised smaller rounds this cycle. Neko's differentiator is the scanner hardware itself: Neko designs its own capture rig rather than routing patients through hospital MRI suites, which lets the company own the capture-to-report pipeline end-to-end. The four-fold valuation jump reflects investor conviction that the vertical will be a real consumer category, not a boutique service.

EconomicTech

Emergent joins the unicorn club at $1.5 billion on a $130M Series C, sixteen months after launch, at $120M ARR

TechCrunch · 2026-07-15 · Source ↗

Bengaluru-based Emergent, an AI coding startup selling an "engineering team in a box" to SMBs, closed a $130 million Series C at a $1.5 billion post-money led by Creaegis, five times its January valuation. The round takes total funding to $230 million.

Bengaluru-based AI coding startup Emergent closed a $130 million Series C at a $1.5 billion post-money valuation Wednesday, entering the unicorn club sixteen months after launch. The round was led by Creaegis, a private-equity firm. New investors MNI Ventures-Claypond and Sentinel Global joined, and existing backers Khosla Ventures, SoftBank Vision Fund 2, Lightspeed, and Y Combinator followed on. Total funding stands at $230 million. Emergent's valuation is up five-fold from the $300 million post-money it carried at its Series B in January 2026.

The company sells software-development-as-a-service to entrepreneurs and small-to-mid-sized businesses that need working code but do not have engineering teams. The pitch, per CEO Mukund Jha: "You're basically getting an engineering team in a box." Under the hood the product is a stack of coding agents that handle programming, deployment, hosting, testing, and debugging, driven from a natural-language interface. Emergent targets trucking companies, factories, construction businesses, property managers, and other non-technical operators who need custom software but cannot spec out or run a build.

"You're basically getting an engineering team in a box."— Mukund Jha, Emergent CEO

The numbers behind the round

Emergent runs at $120 million annualized revenue run-rate, up 70 percent in the past four months. The company has more than 200,000 paying customers. Roughly a third of revenue comes from North America, a third from Europe, and the rest from other markets, with the bulk of engineering staff in Bengaluru. The company employs about 200 people and plans to expand its San Francisco office by 30 to 40 heads by year-end. Brothers Mukund Jha (CEO) and Madhav Jha (CTO) founded the company in June 2025.

Why the Jha bet lands

The unicorn milestone marks the fifth new AI coding unicorn in 2026 alongside Cursor, Cognition, Replit, and Codeium's spinout Devin Local. What sets Emergent apart is customer profile: it does not sell to engineers who want a better IDE. It sells to owners of small businesses that need software but have no in-house build capacity. The category name Emergent uses for that customer is "vibe-coding SMBs" — non-technical operators who describe intent and expect working systems. If that customer profile monetizes, Emergent's addressable market is meaningfully larger than the developer-tool category.

Economic

Bricklaying-robot maker Monumental raises $32M Series B led by Khosla, targets US launch this year

Fortune · 2026-07-15 · Source ↗

Amsterdam-based Monumental, founded by two ex-Palantir engineers, runs 150 autonomous bricklaying robots for European general contractors and now wants to bring the fleet across the Atlantic. Khosla Ventures leads; Hummingbird and Plural follow on.

Amsterdam-based Monumental closed a $32 million Series B Wednesday led by Khosla Ventures. Existing investors Hummingbird and Plural participated. The company was founded in 2021 by Salar al Khafaji (CEO) and Sebastiaan Visser (CTO), the former co-founders of Silk, an analytics company acquired by Palantir in 2016. Monumental deploys autonomous robots that lay brick and mortar with millimeter precision under supervision from human operators.

The company runs a fleet of more than 150 robots as an autonomous bricklaying subcontractor for European general contractors. The economic model is subcontractor-style: Monumental deploys robots to a construction site, its human operators supervise, and the general contractor pays per square meter of wall completed. The Wednesday funding funds a US market entry this year, scale-up of the robot fleet across Europe, and expansion of the range of construction tasks the robots handle beyond bricklaying alone.

"He sold his last company to Palantir. Now he's betting $32 million that robots can fix construction's labor crisis."— Fortune, headline

The bet

Construction is one of the last major industries where labor-driven productivity has failed to improve for decades. The bear case on construction robotics is that job sites are unstructured environments where perception and manipulation are hard, and that the industry's contractor structure resists integration of new tools. Monumental's counter is that its robots ship completed walls, not partial automation, and get paid on the delivery. The strategy has worked in Europe. Whether US general contractors adopt at the same pace is the open question for the Series B thesis.

Economic

InstaLILY AI raises $60M Series B led by Energize Capital, with Insight Partners and Home Depot Ventures joining

AlleyWatch · 2026-07-15 · Source ↗

InstaLILY AI deploys autonomous forward-deployed engineers that learn a customer's business operations, build the software the operation needs, and go live in days. The Wednesday round takes total funding to roughly $100 million.

InstaLILY AI closed a $60 million Series B Wednesday led by Energize Capital, with Insight Partners, Home Depot Ventures, and United Rentals joining. Total equity funding to date reaches roughly $100 million. Founded in 2023 by Amit Shah and Sumantro Das, InstaLILY sells what it calls "autonomous forward-deployed engineers" — coding-and-integration agents that learn a customer's business operations, build the software the operation needs, and go live within days. The pitch is close to Palantir's forward-deployed-engineer model, but the FDE is an agent, not a headcount.

The Home Depot and United Rentals investor participation is the tell for InstaLILY's early market: the company sells into industrials and physical-service businesses where custom operational software is expensive to source through traditional channels. A retail chain building an inventory forecasting overlay, a rental company building a maintenance-schedule optimizer, or a distribution business building a route-planning app has historically had to choose between long consulting engagements and off-the-shelf tools that don't fit. InstaLILY's product produces a fit-to-purpose system in days. Whether the maintenance and change-management costs on autonomous-agent-built software remain low as customers accumulate them is the open question.

EconomicApplication

Justin McLeod's Overtone raises $18M Seed for AI voice-driven matchmaking, with Match Group on the cap table

AlleyWatch · 2026-07-15 · Source ↗

Hinge founder Justin McLeod's new startup Overtone raised $18 million in seed funding from FirstMark, Pace Capital and Match Group for an AI-powered matchmaker that uses voice interviews and relationship-science signals to curate introductions.

Overtone, founded in 2025 by Justin McLeod, closed an $18 million Seed round Wednesday. FirstMark, Pace Capital, and Match Group co-led. McLeod previously founded and led Hinge for 13 years; Match Group acquired Hinge in 2019 and made it the fastest-growing dating app in the company's portfolio. Overtone is McLeod's follow-on: an AI-driven matchmaking service that runs voice interviews with users, extracts compatibility and relationship-goal signals grounded in relationship science, and then presents curated introductions instead of the endless-swipe feed model that Hinge itself popularized.

Match Group's participation is the interesting signal. The parent company of Tinder, Hinge, Match, OkCupid and other core dating brands is investing in an alternative to its own swipe-based products. The internal read on that is that Match sees the swipe-app model as saturated among the Gen Z cohort and sees voice-driven, AI-mediated matchmaking as the plausible next-generation surface. Overtone has not published a product launch date. Whether McLeod ships the voice-first product as the primary interaction, or ships a hybrid voice-plus-app pattern, remains the open UX question.

TechTrend

The AuraStack launch closes the loop: AI agents now design the chips and boards that run AI

The Register · 2026-07-15 · Source ↗

The Cadence launch sits alongside Cadence's May Virtual Engineer for chip design and the June Anthropic Claude Silicon partnership at Synopsys, mapping a stack where LLMs plan and EDA tools execute across every stage of hardware development.

The Cadence AuraStack launch on Wednesday is the third major agentic hardware-design shipment in six weeks. On May 31, Cadence unveiled its "Industry's First Fully Autonomous Virtual Engineer for Chip Design, powered by NVIDIA" — the chip-side equivalent of what AuraStack does at the board and package level. In late June, Synopsys announced its Claude Silicon design partnership with Anthropic, embedding Claude at the RTL analysis and verification stage. AuraStack now fills the PCB-plus-packaging gap.

Read together, the shipments describe a full-stack thesis. The RTL and verification layer runs on Claude and Synopsys's own agents. The chip physical-design layer runs on Cadence Virtual Engineer plus Nvidia GB200 hardware. The board and advanced-packaging layer runs on AuraStack. TSMC sits at the fab-facing edge as the co-development partner for advanced packaging convergence. The three EDA agents share a common architecture: an LLM plans, existing EDA tools execute the physical work, and the LLM iterates until the design converges.

"Cadence's AuraStack agent melds AI with HPC to speed PCB and advanced packaging design."— The Register, on the July 15 launch

The self-reinforcing loop

The industry consequence is a self-reinforcing feedback loop. AI chip design agents like AuraStack run on Nvidia GB200 accelerators. Those same accelerators are being packaged with HBM4 in CoWoS-L configurations that AuraStack is being used to help design. Nvidia Blackwell Ultra silicon planned for 2027 is being designed by agents running on Blackwell. TSMC's advanced packaging capacity is sold out through 2027, meaning the pressure to compress design cycles is not going away. The trend Bloomberg's John Pierre-Jean flagged in his June column — that hyperscaler-scale AI capital expenditure creates the demand for AI-driven EDA that then compresses the design cycle that unlocks more AI capex — is now the operational reality.