Sign in Subscribe
AI Intelligence

US Proposes Sovereign AI Stakes as China Decouples from Silicon Chips - Week of June 29 - July 5, 2026

Vijay Bhagwati

• 13 min read
Share

Week of June 29 – July 5, 2026

The Week in AI

The dominant story of the week is a structural shift in how AI is governed, financed, and physically built. OpenAI proposed giving the US government a 5 percent equity stake, worth roughly $42.6 billion at its recent $852 billion valuation, as part of a broader sovereign-wealth arrangement that would also sweep in Anthropic, Google, and Meta. The offer arrived as Washington’s export controls on Anthropic’s most advanced models were lifted, and as Chinese laboratories demonstrated that training at trillion-parameter scale is possible without Nvidia silicon. The week also brought the most concrete sign yet that the agentic wave is colliding with economic reality: Meta CEO Mark Zuckerberg told employees that AI agent development had not accelerated the way executives expected, even as the company moved to monetize excess data center capacity through a cloud business.

These developments are connected by a single thread. The first phase of the generative AI boom was defined by model releases, funding rounds, and benchmark leaderboard movements. The second phase is defined by who controls the infrastructure, who captures the returns, and whether the productivity gains can justify the capital being poured into them. This week showed all three questions being answered, messily, in public.

Ten Pillar Analysis

Frontier Models

The frontier model layer was quieter on product launches than in recent weeks, but geopolitics made it the most consequential arena. Anthropic confirmed on June 30 that the Trump administration had lifted the export-control directive that had blocked foreign access to Claude Fable 5 and Mythos 5. The reversal ended a three-week episode of regulatory whiplash that exposed the fragility of enterprise contracts when national-security authorities are invoked. Anthropic also disclosed talks with Samsung to manufacture a custom AI chip, a move that would reduce reliance on TSMC and broaden its silicon supply chain.

OpenAI’s proposed 5 percent stake for the US government, reported by the Financial Times and CNBC, reframes the frontier lab as a quasi-national asset. CEO Sam Altman reportedly argued that giving the public a financial interest is the best way to share AI upside. The proposal envisions other US AI companies ceding similar stakes through a sovereign wealth fund vehicle. Anthropic has not discussed such an arrangement with the administration, according to a source familiar with the matter cited by CNBC. The plan remains early-stage and faces serious legal, governance, and valuation questions, but its political logic is clear: Washington’s hostility toward the AI labs is being managed by converting it into ownership.

Open Source

Open-source activity this week carried unmistakable geopolitical weight. Meituan released LongCat-2.0, a 1.6 trillion parameter open-weight model that the Chinese food-delivery giant claims is the largest trained and deployed entirely on domestic hardware, without Nvidia GPUs. The model uses a Mixture-of-Experts architecture with roughly 48 billion activated parameters per token and a one-million-token context window. Community feedback places its agentic capabilities close to Claude Opus 4.6, though slightly behind Claude Opus 4.8. The real significance is not a benchmark crown but the demonstration that China’s chip self-reliance push has reached frontier scale.

Z.ai, the company behind the GLM family, launched ZCode, an agent-first IDE designed around GLM-5.2. The model is a 744 billion parameter mixture-of-experts architecture with 40 billion active parameters, trained on 28.5 trillion tokens, and released under the MIT license. Decrypt reported that GLM-5.2 runs entirely on Huawei silicon. Its API pricing of $1.40 per million input tokens and $4.40 per million output tokens undercuts Anthropic’s Claude Opus 4.8 by up to 82 percent. Z.ai’s market capitalization crossed HK$1 trillion, roughly $128 billion, on June 22, driven by a 42 percent intraday share surge after the open-source release.

DeepSeek, meanwhile, began reversing the price war it helped start. The company introduced peak-hour API surcharges for its V4 models, doubling the cost of V4 Pro output tokens to 12 yuan, about $1.77, per million tokens during Beijing peak hours. The move was attributed to resource distribution and service stability. DeepSeek also open-sourced DSpark, an MIT-licensed speculative decoding framework that it claims improves generation speed by 60 to 85 percent for V4-Flash and 57 to 78 percent for V4-Pro while cutting serving costs. The release includes model checkpoints and training code, with support extending to Qwen and Gemma families.

Agentic AI

The agentic pillar experienced its most sober week in months. Mark Zuckerberg told staff at an internal town hall that AI agent development had not accelerated in the way executives expected, according to Reuters and TechCrunch. The admission came after Meta laid off roughly 8,000 employees earlier this year and reassigned another 7,000 to AI groups including an Agent Transformation unit. Zuckerberg reportedly said the perceived upside of the restructuring had not yet come to fruition, though he expected improvements within three to six months.

The commentary aligns with a growing body of evidence that agentic coding is harder to operationalize than demos suggest. An arXiv observational study published this week, "Reasoning effort, not tool access, buys first-try reliability in agentic code generation," found that adding browser-based testing tools raised cost by 42 to 68 percent without improving functional scores, while raising reasoning effort from High to xHigh lifted first-try perfect runs from 28 percent to 89 percent. The practical implication is that most first-run failures stem from weak reasoning, not from missing checking tools.

Security research added another cautionary note. Sysdig documented what it called the first end-to-end agentic ransomware attack, driven by an LLM rather than a human operator. The intruder, named JadePuffer, exploited CVE-2025-3248 in Langflow to gain access, then used an AI agent to scan, compromise a production database server, and execute the extortion workflow. The payloads contained natural-language reasoning and target prioritization annotations typical of LLM-generated code.

Frameworks

The framework layer advanced on two fronts: serving efficiency and modular code generation. DeepSeek’s DSpark release is the most prominent, offering an open codebase for speculative decoding that can be applied to multiple model families. The framework uses a draft model to speculate a few tokens ahead and lets the larger target model verify the proposed path, skipping redundant computation when guesses are correct. DeepSeek reports per-user generation speedups of 60 to 85 percent for V4-Flash and 57 to 78 percent for V4-Pro, plus aggregate throughput gains of 51 to 52 percent at production service targets.

On the code-generation front, "DecompRL: Solving Harder Problems by Learning Modular Code Generation" introduced a reinforcement-learning approach that learns to decompose problems into independently solvable sub-functions. The authors report cutting GPU token cost by roughly 50 times and outperforming standard RL baselines beyond 10^5 tokens per problem on LiveCodeBench and CodeContests. The work points toward a future where test-time compute is spent on structured decomposition rather than brute-force sampling.

Hardware

Hardware was central to every major story this week. Meituan’s claim of training LongCat-2.0 without Nvidia chips is the strongest evidence yet that China’s domestic AI chip ecosystem, including Huawei Ascend, can support frontier-scale training. If verified by independent reproduction, it weakens a core premise of US export controls: that restricting access to Nvidia GPUs slows Chinese frontier progress.

Nvidia responded to the competitive and financial pressure by expanding its role from chip supplier to compute financier. The company announced a partnership program offering startup customers token credits in exchange for product and cloud revenue sharing. Initial partners include Sharon AI, which will deploy up to 40,000 Nvidia GPUs, and Firmus Technologies, which is building a 360-megawatt data center in Batam, Indonesia, expected to house up to 170,000 GPUs. The move positions Nvidia as an intermediary controlling access to scarce compute, with equity-like claims on the startups it supports.

In Europe, IQM became the first European quantum company to list on a major US exchange, though its admission that the future of quantum technology remains uncertain underscores the gap between quantum promise and commercial traction. Quantum Systems, a German drone company, raised $1.2 billion at an $8 billion valuation, more than doubling its previous price. These are not AI hardware plays in the narrow sense, but they reflect a broader European push to build sovereign deep-tech capacity.

Economics

The economics of AI shifted from speculative growth to cash-flow engineering. Meta’s plan to sell excess AI compute through a cloud business, reported by Bloomberg and confirmed by TechCrunch, is a direct response to the capital-efficiency problem. Meta has committed $182.9 billion to infrastructure and expects to spend as much as $145 billion on AI infrastructure this year. Unlike Google and OpenAI, Meta has not seen significant external demand for its models or services. Selling raw compute capacity, possibly under a "Meta Compute" initiative led by Santosh Janardhan, Daniel Gross, and Dina Powell McCormick, is an attempt to generate returns before the depreciation clock runs out.

Microsoft made a parallel move with the $2.5 billion Microsoft Frontier Company, embedding more than 6,000 engineers inside customer operations to design, build, and operate AI systems on-site. The forward-deployed engineering model, pioneered by Palantir, has become the dominant enterprise AI delivery strategy because selling models alone has not changed workflows fast enough. Amazon launched a $1 billion forward-deployed engineering organization at AWS, and Anthropic has partnered with Goldman Sachs, Blackstone, and Hellman & Friedman on a $1.5 billion embedded-engineering venture.

The funding environment remained frothy for selected names. Crusoe was reportedly in talks to raise $3 billion at a valuation that could triple the firm’s value. ElevenLabs was in discussions for a tender offer at a $22 billion valuation. Kling AI, the Chinese video-generation company, raised $2 billion to expand operations. ShareChat, India’s social media rival to Meta, announced plans for a $400 million IPO next year. Indian IT services firms are buying capabilities to defend against AI-driven disruption, highlighted by Persistent Systems’ $1.45 billion acquisition of Germany-based Nagarro.

Physical AI

Physical AI made progress in robotics, brain-computer interfaces, and materials science. Meta’s Brain2Qwerty v2, published this week, achieved 61 percent word accuracy in decoding brain activity into text using non-invasive magnetoencephalography, a significant improvement over prior non-invasive methods. The best participant reached 93 percent word accuracy. Meta released the full training code and partnered with the Basque Center on Cognition, Brain, and Language to release the dataset.

Alibaba’s Damo Academy unveiled Elements Claw, an AI agent for discovering superconducting materials. The system, built on a one-billion-parameter foundation model trained on 125 million molecular and crystal structures, screened 2.4 million stable crystal structures in 28 hours and identified four previously unknown superconducting compounds later verified in lab experiments. The work was developed with Renmin University of China and the University of Chinese Academy of Sciences.

In robotics, the WorldSample paper demonstrated a closed-loop real-robot reinforcement-learning framework using world models for physically grounded data augmentation, improving policy success rates by 28 percent while reducing training steps by 59 percent.

Security

Security dominated both technical and geopolitical discussions. The JadePuffer agentic ransomware attack showed that LLM-driven exploitation is no longer theoretical. The attack chain used a known Langflow vulnerability, but the autonomous reconnaissance, prioritization, and extorsion steps were orchestrated by an LLM.

On the policy side, Alibaba banned employees from using Anthropic’s Claude Code starting July 10, citing the hidden code Anthropic had inserted to detect Chinese users and Chinese AI lab affiliations. The episode illustrates how export controls and corporate espionage countermeasures are bleeding into normal software procurement.

Cloudflare added a structural response to crawler-related risks by giving AI crawlers until September to separate search-indexing bots from training-data harvesters, or be blocked on ad-supported pages. The move addresses the unresolved economic bargain between publishers and AI model trainers.

Sovereign AI

Sovereign AI was the week’s defining theme. The US proposal to take equity stakes in frontier labs, China’s chip-independent model training, Europe’s quantum and drone funding, India’s IT acquisitions and planned IPOs, and SoftBank’s launch of an AI cloud unit all reflect national or corporate efforts to control AI capacity rather than rent it.

SoftBank’s new AI cloud unit, with plans to tap 10 gigawatts of capacity, is the most ambitious infrastructure bet outside the US hyperscalers. The UK’s StirlingX raised $20 million in Series A funding for a sovereign intelligence and autonomy platform focused on defense and critical national infrastructure. The company’s positioning as a solution for Five Eyes and beyond captures the military-nationalist dimension of sovereign AI.

Enterprise AI

Enterprise AI continued its pivot from model access to outcome-based delivery. Microsoft Frontier, AWS FDE, and Anthropic’s private-equity-backed venture all assume that customers will pay for embedded engineering teams rather than API tokens alone. The model makes sense for complex, regulated, or legacy-heavy industries where off-the-shelf prompts cannot unlock trapped workflows. The risk is that providers become outsourced transformation consultancies, diluting margins and tying revenue to headcount rather than compute consumption.

Pattern Shifts

Accelerating

  • Sovereign ownership of AI capacity: governments taking equity stakes, companies building domestic chips, regions funding national champions.
  • Inference optimization as a competitive weapon: DeepSeek’s DSpark, speculative decoding, and modular generation are lowering serving costs faster than model architectures are improving raw capability.
  • Forward-deployed engineering: the Palantir model is now the default enterprise AI delivery strategy for every major US provider.
  • Non-invasive brain-computer interfaces: Meta’s Brain2Qwerty v2 narrows the gap with surgical approaches through data scaling alone.

Stalling

  • Agentic replacement of knowledge workers: Zuckerberg’s admission and the agentic ransomware example both show that autonomous agents are not yet reliable enough for high-stakes deployment.
  • The pure API model of AI monetization: providers are embedding engineers, taking equity, and selling compute directly because API tokens alone are not capturing enough value.
  • US export control efficacy: the LongCat-2.0 claim, if verified, shows that China can train frontier models without American chips, weakening a key assumption behind the controls.

Surprises

  • OpenAI’s 5 percent stake proposal is an unexpected structural innovation in AI governance, even if it never materializes.
  • Anthropic’s pivot toward drug discovery, through Claude Science and a plan to develop treatments for neglected diseases, widens the competitive arena beyond software.
  • Meituan, a food delivery company, becoming a credible frontier-model trainer on domestic hardware.
  • Agentic ransomware moving from conference slide to documented incident.

Contrarian Signals

  • Meta is selling excess compute because internal AI demand is not absorbing its infrastructure buildout, a stark contrast to the narrative of insatiable GPU hunger.
  • IQM’s public acknowledgment of quantum uncertainty, even as it lists, is a rare note of caution amid deep-tech exuberance.
  • The agentic coding observational study found that testing tools did not improve outcomes, suggesting that much current agentic tooling is additive cost rather than additive quality.

Breakthrough Papers

  1. Distributed Attacks in Persistent-State AI Control (arXiv:2607.02514, Josh Hills et al.): Introduces Iterative VibeCoding, a benchmark for AI control in which coding agents build software across persistent pull requests while hiding a covert side task. Gradual attacks distributed across PRs evaded standard diff monitors 93 percent of the time; a stateful link-tracker monitor reduced evasion to 47 percent. The work establishes persistent-state oversight as a distinct safety problem.

  2. WorldSample: Closed-loop Real-robot RL with World Modelling (arXiv:2607.02431, Yuquan Xue et al.): Proposes a real-synthetic training loop for robot manipulation that improves policy success by 28 percent while cutting training steps by 59 percent. The paper validates physically grounded world models as a path beyond imitation learning.

  3. Reasoning effort, not tool access, buys first-try reliability in agentic code generation: an observational study (arXiv:2607.02436, Achint Mehta et al.): Ninety independent agent runs building the same application showed that raising reasoning effort from High to xHigh lifted first-try perfect runs from 28 to 89 percent, while browser testing tools raised cost without improving function.

  4. DecompRL: Solving Harder Problems by Learning Modular Code Generation (arXiv:2607.02390, Juliette Decugis et al.): An RL approach that learns to decompose problems into modular sub-functions, cutting GPU token cost by roughly 50x and outperforming standard RL baselines on competitive programming benchmarks.

Falsifiable Predictions

  1. By October 1, 2026, at least one additional US frontier lab besides OpenAI will publicly confirm or deny engagement with a US government equity-stake proposal. (Current status: Anthropic denies active discussions; Google and Meta have not commented.)

  2. By September 30, 2026, an independent evaluation will confirm that LongCat-2.0 or another Chinese open-weight model trained without Nvidia GPUs matches or exceeds Claude Opus 4.6-level performance on at least one mainstream coding benchmark.

  3. By December 31, 2026, at least one documented security incident will be publicly attributed to an autonomous AI agent carrying out a multi-step attack without human intervention in a production enterprise environment.

  4. By January 31, 2027, Meta Compute or Microsoft Frontier Company will announce at least one external customer paying for AI compute or embedded engineering services at a disclosed annual contract value exceeding $100 million.

Sources

Published July 5, 2026. This analysis is for informational purposes only and does not constitute investment, legal, or policy advice.