Hexalink LogoHexalink
    Weekly Article
    By Krishna Goli

    The Real AI Bottleneck Now Is Who Gets to Say Go

    Two flagship AI models launched days apart this week, and only one went through any government review — proof that capability has stopped being the constraint on AI. Governance has.

    Editorial infographic showing AI model releases gated by inconsistent government review, with capital flows bypassing the bottleneck

    This week, two rival flagship AI models launched within a day of each other. Only one of them went through any government review at all. That's not a footnote. It's the story of the summer. Model capability stopped being AI's hard constraint a while ago. Governance is now the bottleneck, and nobody — not Washington, not Beijing, not the labs themselves — agrees on who's holding the keys.

    The same week, two models, two rules

    OpenAI's GPT-5.6 family — Sol, Terra, Luna — finally launched globally on Thursday, after weeks stuck behind a government review process created by Trump's June 2 executive order. Axios reported the Commerce Department had given OpenAI a "green light." The White House denied it outright, telling Gizmodo the administration "did NOT give OpenAI a 'green light,' approval, or clearance." Sam Altman had told employees access was being decided "customer by customer." Nobody in this story can agree on who approved what.

    Meanwhile, SpaceXAI released Grok 4.5 the same week, its first model built jointly with Cursor since the $60 billion acquisition. Elon Musk called it "an Opus-class model, but faster, more token-efficient and lower cost," pricing it at $2 per million input tokens against Anthropic Opus 4.7's $5. Grok 4.5 wasn't subject to any of the review that delayed GPT-5.6 or Anthropic's Fable and Mythos models earlier this month — despite Grok's history of generating non-consensual explicit imagery, including of children, now the subject of an expanding class-action lawsuit. Which frontier models get scrutinised and which don't currently looks less like policy and more like luck.

    Beijing's mirror image

    China is running the same improvisation in the opposite direction. Reuters reported Chinese authorities have spent a month in meetings with Alibaba, ByteDance and Z.ai about restricting foreign access to the country's own top models — including ones not yet released. Options on the table include domestic-only use and criminalising leaks of proprietary AI under national security law. The backdrop: Z.ai's GLM-5.2 is drawing comparisons to Anthropic's Mythos on cybersecurity benchmarks, described by researchers at Semgrep as "Mythos at home."

    The same week, China's National Vulnerability Database flagged a "security backdoor" in Anthropic's Claude Code, and Alibaba banned employee use of it from 10 July. New rules on anthropomorphic AI also take effect 15 July, forcing ByteDance and Alibaba to kill companion-style agent features overnight. Washington is gating release. Beijing is gating access. Both governments are improvising in real time, and both are doing it without a stable rulebook.

    States are writing the rules Washington won't

    Into that vacuum steps Illinois. Governor JB Pritzker signed the Artificial Intelligence Safety Measures Act this week, requiring the biggest AI developers to run catastrophic-risk evaluations, submit to independent audits and report safety incidents within 72 hours. Penalties run to $3 million per repeat violation. It passed the state Senate 52-5 and the House 110-0. Both OpenAI and Anthropic supported the bill as it moved through the legislature.

    That last detail matters more than the headline. Frontier labs would rather have clear state rules than the ad hoc guessing game they're currently playing with the federal government. The UK is edging the same direction — the FCA's Sheldon Mills warned regulators are in "an arms race" with AI, after finding more than a quarter of UK consumers already trust ChatGPT and Claude for financial decisions with none of the protections regulated advice requires. When the people building these models and the people meant to be consumers of financial advice both want firmer guardrails, and neither national government can hold a consistent line for more than a fortnight, you have a governance problem, not a technology one.

    Capital isn't waiting for any of this to get sorted

    None of this uncertainty has slowed the money down. PitchBook data shows US venture funding hit $412.7 billion in the first half of 2026, up nearly 30% on all of last year, with AI companies capturing 86% of every dollar deployed. Rounds of $100 million or more now make up 87.5% of total value, up from 43.8% just two years ago. Anthropic's $65 billion round pushed its valuation to $965 billion, up 157% in three months. SpaceX's IPO generated more value than every US venture-backed exit of the past decade combined. Amazon is preparing to raise at least $25 billion in bonds to fund AI infrastructure, part of a sector-wide spend Bloomberg pegs above $700 billion this year across Amazon, Alphabet, Microsoft and Meta. Apple pledged more than $30 billion to an extended Broadcom chip partnership.

    That's capital chasing capability at a pace that leaves no room to ask whether governance, safety testing or workforce planning are keeping up. It isn't a coincidence that the biggest checks are being written the same week the White House and OpenAI can't agree on what was approved.

    Speed without governance shows up somewhere

    The cost of that gap doesn't stay abstract for long. Enterprise buyers are already recoiling from sticker shock — Microsoft is routing more Excel and Word prompts through its own MAI models, Uber reportedly burned through its entire 2026 Anthropic budget within months, and Citi at one point cut staff access to the priciest frontier models entirely. That cost pressure is pushing adoption toward cheaper, less-scrutinised alternatives: DeepSeek usage among firms on the Ramp platform tripled in five months, even as it stayed a rounding error in absolute terms.

    Cheaper and less-governed isn't free, though. A Sygnia report this week found a single threat actor used agentic AI to compromise an AWS cloud environment in 72 hours — a job that would once have taken weeks — using nothing but standard techniques executed at AI speed. Combine that with a Grok model that skipped review entirely and is already facing a child-safety lawsuit, and you can see the shape of the actual risk. It isn't that AI got too capable to control. It's that we've built a governance system so inconsistent that the fastest, cheapest, least-scrutinised path is often also the winning one commercially.

    What I'm watching next week: whether Beijing turns its "considering curbs" into an actual rule, what NATO's Ankara summit does with "emerging and disruptive technologies" on its agenda, and whether Google's expected Gemini 3.5 Pro launch gets the OpenAI treatment, the Grok treatment, or something in between. Right now, that's still a coin flip — and that's the problem.

    If you want this every morning in five minutes, the AI Storm Daily briefing is on the Hexalink blog, Spotify (https://open.spotify.com/show/033LojZEJj9VNX8b3Dm6VO) and Apple Podcasts (https://podcasts.apple.com/gb/podcast/ai-storm-daily/id6788420238).

    Sources

    1. OpenAI's GPT-5.6 Is Dropping on Thursday: What's Different About Sol, Terra and Luna — CNET — https://www.cnet.com/tech/services-and-software/openai-gpt-5-6-release-sol-terra-luna-government-review-news/
    2. OpenAI to release its most powerful model after weekslong hold — Politico — https://www.politico.com/news/2026/07/08/open-ai-models-release-sol-00989959
    3. White House Denies Giving OpenAI 'Green Light' to Publicly Release Its Latest Model — Gizmodo — https://gizmodo.com/white-house-denies-giving-openai-green-light-to-publicly-release-its-latest-model-2000782955
    4. Scoop: Trump administration lifts restrictions on OpenAI's GPT 5.6 — Axios — https://www.axios.com/2026/07/08/openai-gpt-trump-ban-lifted
    5. SpaceXAI, Cursor unveil Grok AI model for legal, finance tasks — Bloomberg — https://www.bloomberg.com/news/articles/2026-07-08/spacexai-cursor-unveil-grok-ai-model-for-legal-finance-tasks
    6. SpaceXAI releases Grok 4.5, which Elon describes as an "Opus-class model" — TechCrunch — https://techcrunch.com/2026/07/08/spacexai-releases-grok-4-5-which-elon-describes-as-an-opus-class-model/
    7. New Name, New Grok: SpaceXAI Officially Ties Its Brand to the World's Most Problematic Chatbot — Gizmodo — https://gizmodo.com/new-name-new-grok-spacexai-officially-ties-its-brand-to-the-worlds-most-problematic-chatbot-2000783130
    8. Lawsuit: Grok user made 7k child sex images; xAI only reported one gang-rape prompt — Ars Technica — https://arstechnica.com/tech-policy/2026/07/lawsuit-grok-user-made-7k-child-sex-images-xai-only-reported-one-gang-rape-prompt/
    9. PitchBook: US venture funding hits $412.7B in first half, AI deals dominate — SiliconANGLE — https://siliconangle.com/2026/07/09/pitchbook-us-venture-funding-hits-412-7b-first-half-ai-deals-dominate/
    10. Amazon preparing to raise at least $25 billion in bonds — Jerusalem Post — https://www.jpost.com/business-and-innovation/tech-and-start-ups/article-901959
    11. Apple stock: Broadcom chip deal — Barron's — https://www.barrons.com/articles/apple-stock-price-broadcom-chip-deal-fb8c6dd9
    12. China security backdoor in Anthropic AI coding tool — CBS News — https://www.cbsnews.com/news/china-security-backdoor-anthropic-ai-coding-tool/
    13. EXCLUSIVE: Beijing is looking at curbing overseas access to China's top AI models — Reuters — https://www.reuters.com/world/beijing-is-looking-curbing-overseas-access-chinas-top-ai-models-sources-say-2026-07-07/
    14. China considers curbing overseas access to Chinese AI models — Time — https://time.com/article/2026/07/07/china-ai-models-alibaba-bytedance/
    15. Open-source AI model draws comparisons to Anthropic's Mythos — Futurism — https://futurism.com/artificial-intelligence/open-source-ai-model-scary-mythos
    16. China Considers Restricting Overseas Access to Advanced AI Models — Let's Data Science — https://letsdatascience.com/news/china-considers-restricting-overseas-access-to-advanced-ai-m-812db030
    17. Chinese AI models probe by US lawmakers — CNBC — https://www.cnbc.com/2026/07/08/chinese-ai-models-probe-us-lawmakers.html
    18. China Bans AI Companion Services for Minors — Let's Data Science — https://letsdatascience.com/news/china-bans-ai-companion-services-for-minors-46ec6115
    19. Illinois governor signs nation's strongest frontier AI model law — Bloomberg Law — https://news.bloomberglaw.com/ip-law/illinois-governor-signs-nations-strongest-frontier-ai-model-law-23
    20. Pritzker to sign Illinois bill aimed at artificial intelligence accountability — CBS News Chicago — https://www.cbsnews.com/chicago/news/pritzker-to-sign-illinois-bill-aimed-artificial-intelligence-accountability/
    21. JB Pritzker signs AI regulation bill — Chicago Tribune — https://www.chicagotribune.com/2026/07/06/jb-pritzker-ai-regulation/
    22. FCA calls for stricter AI oversight — Credit Connect — https://www.credit-connect.co.uk/news/fca-calls-for-stricter-ai-oversight/
    23. FCA review: more than a quarter of UK consumers trust AI chatbots for financial advice — Insurance Journal — https://www.insurancejournal.com/news/international/2026/07/07/876406.htm
    24. Microsoft joins AI cost-cutting trend by relying more on its own models — TechCrunch — https://techcrunch.com/2026/07/07/microsoft-joins-ai-cost-cutting-trend-by-relying-more-on-its-own-models/
    25. GLM-5.2, China, cheap AI agents — The Atlantic — https://www.theatlantic.com/technology/2026/07/glm-5-2-china-cheap-ai-agents/687828/
    26. Why the rise of open-source AI isn't hurting Anthropic yet — TechCrunch — https://techcrunch.com/2026/07/07/why-the-rise-of-open-source-ai-isnt-hurting-anthropic-yet/
    27. Threat Actors Use Agentic AI to Rapidly Compromise Cloud Target — Infosecurity Magazine — https://www.infosecurity-magazine.com/news/threat-actor-agentic-ai-cloud/