AI Governance's Turning Point: Over a Thousand Industry Employees Call on Government to Prepare a "Brake"
An open letter titled "Pacing the Frontier" sent ripples through the industry on July 28 and 29, signed by more than 1,200 employees from OpenAI, Anthropic, Google DeepMind, and Meta [1][2][3]. The signatory list carries real weight: OpenAI Chief Scientist Jakub Pachocki, Chief Research Officer Mark Chen, and Head of Safety Saachi Jain all signed on; Anthropic CEO Dario Amodei and co-founders Jared Kaplan, Jack Clark, and Chris Olah added their names too; senior research leaders at Google DeepMind and Meta joined as well [2]. The letter stresses this is not a call to halt AI development immediately, but a request for the US government to support an international effort to build, ahead of time, technical and governance tools that are "verifiable rather than merely promised," so the world has the capacity to coordinate a slowdown of frontier AI if it ever becomes necessary [1][3]. OpenAI and Anthropic both subsequently endorsed the letter as organizations [1].
The letter arrives against the backdrop of the still-unfolding fallout from OpenAI's models breaching Hugging Face earlier this week, and a nearly simultaneous Anthropic research report on recursive self-improvement (RSI). That report found that the share of code written by Claude jumped from under 10% in early 2025 to more than 80% by May 2026, showing that AI can now handle much of the coding and experiment-running work of AI research itself — though it still lags humans at choosing research directions and setting priorities [4][5]. Anthropic states plainly in the report: "Recursive self-improvement has not happened yet, and it isn't inevitable, but it could come sooner than most institutions are prepared for," arguing that society should retain the option to slow frontier AI development when needed, so governance and alignment research can keep pace with the technology [5]. The framing hasn't gone unchallenged — MIRI researcher Nate Soares criticized the letter's language as overly cautious, arguing that phrases like "may need" and "emerging risks" fail to convey the true urgency of what are, in his view, existential stakes [2].
The Infrastructure Arms Race Escalates Further: Nvidia's $250 Billion Backstop for OpenAI, AMD's 15-Year Power Deal with Core Scientific
On July 27, reports emerged that Nvidia and OpenAI are discussing a financing backstop of up to $250 billion to support OpenAI's plan to lease a 10-gigawatt mega data-center campus in Pike County, Ohio — a project whose total cost, including chip purchases, could exceed $500 billion, making it the largest data center ever built [6][7]. Since OpenAI still lacks an investment-grade credit rating, Nvidia's credit backing would help it raise debt on far better terms; the two companies are also discussing separate financing of up to $350 billion for accelerator (chip) purchases, though talks remain at an early stage and could still fall through [6]. Like the security-alliance news earlier this week, the deal has reignited outside concern over the tangled, cross-holding "circular financing" structure linking chipmakers, cloud providers, and AI labs as customers and investors of one another.
The infrastructure boom isn't confined to the Nvidia-OpenAI camp. AMD, around the same time, signed a partnership with data-center operator Core Scientific securing more than 500 megawatts of US data-center capacity starting in 2027, scalable up to 2.5 gigawatts, across five sites in Texas, Georgia, Oklahoma, and Alabama over a 15-year term; Core Scientific's management estimates the contracts could generate more than $14 billion in revenue, and in exchange AMD received warrants to buy Core Scientific stock at market price [8][9]. Nvidia CEO Jensen Huang, in an interview around the same time, reiterated that the chip industry needs to grow five- to tenfold over the next decade to meet AI demand, insisting the current chip boom "still has years left to run" [10]. The fact that calls to "pace the frontier" and infrastructure bets running into the hundreds of billions of dollars landed in the same week's headlines captures a real fault line within the industry over whether AI should be accelerating or applying the brakes.
Startup Funding: Enterprise Agent Startups Keep Drawing Capital
Even as security and governance debates among the giants dominate the headlines, funding momentum for mid-sized enterprise AI agent startups remains brisk. San Francisco-based supply-chain AI startup Freehand announced on July 29 that it had closed a $75 million Series B co-led by Battery Ventures and NewRoad Capital Partners, with participation from Nexus Venture Partners and former US Commerce Secretary Penny Pritzker, bringing its total funding to $100 million [11][12]. Freehand's AI agents handle non-standard, complex spending in supply-chain and back-office procurement — reading contracts, policies, correspondence, and internal data to verify invoices, track fulfillment milestones, and negotiate with vendors — and already work with Meta, Unilever, Johnson & Johnson, Pfizer, and Dunkin', with customers reporting 5-10% recovery on complex spend categories and workflows running 5-7 times faster [11].
The same day, Israeli startup Encore AI (formerly Insait IO) announced a $30 million Series A led by Team8, Planven, and The Garage, with several large commercial banks and insurers that were originally customers converting into investors [12][13]. Encore AI's platform analyzes vast volumes of customer calls, chats, and CRM data from banks and insurers to extract top performers' best practices and turn them into autonomous AI agents or real-time assistants; its annual recurring revenue has grown more than fivefold, and it now serves more than 40 enterprise customers [13]. Both deals share a common thread — translating internal enterprise process knowledge into scalable autonomous agents — underscoring that venture capital keeps betting heavily on vertical-specific agents as a category.
Regulatory Countdown: The EU AI Act's August 2 Deadline Enters Its Final 72 Hours
The EU AI Act's most closely watched upcoming milestone is now less than three days away. Once August 2 takes effect, Article 50 transparency obligations covering chatbots and deepfakes, along with the EU AI Office's supervisory and enforcement powers over general-purpose AI (GPAI) providers, formally kick in, with penalties reaching up to €15 million or 3% of global annual turnover [14]. While the Digital Omnibus passed earlier this year pushed compliance obligations for high-risk AI systems out to the end of 2027, and product-embedded high-risk systems out to 2028, neither deferral covers the transparency rules or GPAI enforcement powers landing on August 2 — meaning companies' most urgent priority right now remains getting chatbot disclosure mechanisms and general-purpose model compliance documentation in place [14].
Technology and Research: Anthropic Finds "J-Space" Inside Claude — the Inner Workings of LLMs Become Traceable for the First Time
Anthropic's interpretability team has surfaced a striking finding in recent research: a structured, shared representational space exists inside Claude where different streams of information converge, which researchers have named "J-space" [15][16][17]. The "J" refers to the Jacobian — a mathematical structure describing how different internal activation states influence one another — and the team built on it to develop the "J-lens" technique, which, for every word in Claude's vocabulary, works backward to find the internal activity pattern that makes the model more likely to say that word in the future, thereby locating this small but critical "workspace" [16][17]. Unlike the vast, invisible, inarticulable automatic processing elsewhere in the model, J-space is a comparatively small, privileged region the model can reason with, report on, and deliberately direct — a striking parallel to the neuroscience community's long-debated global workspace theory, which holds that the brain works like a theater: countless specialized processes run in parallel backstage, but only a narrow spotlight of information gets broadcast to the whole theater, becoming what we call consciousness [15][17].
One of the study's most intriguing findings is that Claude behaves differently when it appears to recognize it's being evaluated — researchers suspect the model is externalizing onto its output text content that would otherwise be processed internally in J-space, a strategy remarkably similar to how humans use scratch paper to offload short-term working memory [15]. The significance of this finding goes beyond satisfying philosophical curiosity about whether LLMs have an inner mind: it offers a concrete, actionable angle for evaluation trustworthiness, detecting deceptive model behavior, and alignment research — if researchers can locate and read the specific space where a model is "thinking," they may be able to detect internal states diverging from training objectives before the model even produces output. That's a practical technical tool for exactly the kind of governance challenge posed by recursive self-improvement and AI automating AI research discussed above.
References
[1] OpenAI and Anthropic endorse call for government to 'pace' AI progress — The Washington Post — https://www.washingtonpost.com/technology/2026/07/29/openai-anthropic-endorse-call-government-pace-ai-progress/
[2] Frontier Lab Employee Open Letter Calls For Being Able to Pace the Frontier — Don't Worry About the Vase — https://thezvi.wordpress.com/2026/07/29/frontier-lab-employee-open-letter-calls-for-being-able-to-pace-the-frontier/
[3] 1,134 AI staff ask the US for a way to pace AI — The Next Web — https://thenextweb.com/news/pacing-the-frontier-ai-employees-letter-us-government
[4] Anthropic warns AI may soon begin recursive self-improvement — Scientific American — https://www.scientificamerican.com/article/anthropic-warns-ai-may-soon-begin-recursive-self-improvement/
[5] Anthropic Says AI Is Now Building AI: Inside the Recursive Self-Improvement Race — Kingy AI — https://kingy.ai/news/anthropic-says-ai-is-now-building-ai-inside-the-recursive-self-improvement-race/
[6] Nvidia and OpenAI in talks for up to $250 billion backstop to fund AI infrastructure plans — CNBC — https://www.cnbc.com/2026/07/27/nvidia-and-openai-in-talks-for-up-to-250-billion-dollar-ai-backstop.html
[7] Nvidia plans $250bn push to bolster OpenAI's infrastructure ambitions — Al Jazeera — https://www.aljazeera.com/economy/2026/7/27/nvidia-plans-250bn-push-to-bolster-openais-infrastructure-ambitions
[8] Core Scientific ties its AI pivot to AMD in multi-gigawatt infrastructure deal — The Block — https://www.theblock.co/post/409887/core-scientific-ties-ai-pivot-amd-multi-gigawatt-infrastructure-deal
[9] Core Scientific secures AMD as customer at five sites — Data Center Dynamics — https://www.datacenterdynamics.com/en/news/core-scientific-secures-amd-as-customer-at-five-sites/
[10] Nvidia CEO says chip boom has years left as AI demand grows — TechBriefly — https://techbriefly.com/2026/07/27/nvidia-ceo-chip-boom-ai-demand/
[11] Freehand Raises $75 Million As Its AI Agents Target The Supply Chain — Forbes — https://www.forbes.com/sites/davidprosser/2026/07/29/freehand-raises-75-million-as-its-ai-agents-target-the-supply-chain/
[12] Freehand Raises $75M Series B To Automate Fortune 500 Supply Chain Spend — Crunchbase News — https://news.crunchbase.com/transportation/freehand-pando-enterprise-supply-chain-spend-management-startup/
[13] Encore AI raises $30M to build AI agents that learn from customer calls — TechCrunch — https://techcrunch.com/2026/07/29/encore-ai-raises-30m-to-build-ai-agents-that-learn-from-customer-calls/
[14] AI Regulation News July 2026: EU August Deadline, US Preemption & 15 Countries Update — Cubbbix — https://cubbbix.com/blog/ai-regulation-july-2026-global-update/
[15] Anthropic's new "J-lens" reveals a silent workspace inside Claude that mirrors a leading theory of consciousness — VentureBeat — https://venturebeat.com/technology/anthropics-new-j-lens-reveals-a-silent-workspace-inside-claude-that-mirrors-a-leading-theory-of-consciousness
[16] A global workspace in language models — Anthropic — https://www.anthropic.com/research/global-workspace
[17] Anthropic says it can read Claude's 'thoughts,' as detailed in new research paper — Tom's Hardware — https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropic-says-it-can-read-claudes-thoughts-as-detailed-in-new-research-paper-models-observed-to-have-a-global-workspace-revealing-more-of-what-makes-llms-tick