Washington Gatekept a Model Release. Google Lost a Nobel Laureate. OpenAI Built Its Own Chip.

Two weeks ago, the Commerce Department export-controlled Anthropic's frontier models. This column asked whether that precedent would expand. On Friday, it expanded. The White House asked OpenAI to release GPT-5.6 Sol to a small group of government-approved partners rather than conducting a standard commercial launch. OpenAI complied, publicly, while simultaneously publishing a statement insisting this arrangement should not become the long-term default. In the same seven-day window, Google lost the co-author of the Transformer paper and a Nobel Prize-winning protein researcher to OpenAI and Anthropic, respectively, wiping roughly $240 billion from Alphabet's market capitalization in a single trading session. OpenAI unveiled its first custom silicon. Five intelligence agencies issued a joint warning that AI-enhanced cyberattacks will arrive within months. SpaceX committed $60 billion to acquire the most popular AI coding tool on Earth. Meta launched a $299 pair of AI glasses designed for people who find $379 too expensive for a computer on their face. And SK Hynix, a company that nearly went bankrupt in the 2000s, quietly surpassed Samsung Electronics to become South Korea's most valuable firm for the first time since the millennium, because the chips it bet on in 2012 turned out to be the ones the AI era actually needs.

Nine weeks into this column, a pattern has crystallized that was only emerging in the early installments: the decisions that matter are no longer being made by engineers. They are being made by trade officials, intelligence directors, portfolio managers, and acquisition committees. The technical frontier continues to advance, but the governance layer now determines who gets to touch it, when, and under what conditions. Here are the seven things that defined the week.

1. The White House Told OpenAI Who Could Use GPT-5.6

On June 26, OpenAI launched GPT-5.6 as a three-model suite: Sol, the flagship; Terra, a mid-tier model offering GPT-5.5-level performance at half the cost ($2.50 per million input tokens, $15 per million output); and Luna, the lightest and cheapest ($1 per million input, $6 per million output). Sol introduces two new capabilities. The first is a “max” reasoning mode that gives the model more time to decompose complex problems. The second is “ultra” mode, which spawns multiple sub-agents to parallelize work. Sol is priced at $5 per million input tokens and $30 per million output tokens.

The technical announcement would have been the story in any other month. Not this week.

The story was the release mechanism. The White House asked OpenAI to limit the initial launch to approximately 20 government-approved partners whose identities were shared with the administration. OpenAI CEO Sam Altman disclosed the arrangement in an internal memo that surfaced through industry reporting, calling it a temporary measure to “safely navigate a strange moment.” The company released a public statement emphasizing that the government approval process should not become the default for future model releases. “We believe in broad access,” the statement read, a sentence that only needs to be written when access is being narrowed.

The administration's rationale mirrors its Anthropic intervention: Sol's cybersecurity capabilities, specifically its ability to identify and analyze software vulnerabilities, sit near the threshold that triggered the Fable 5 shutdown. OpenAI's own Preparedness Framework rates GPT-5.5 as “High” on cyber capability. Sol exceeds that baseline. The company argues that Sol is significantly better at helping defenders find and fix vulnerabilities than at conducting end-to-end offensive attacks, a distinction that amounts to saying the knife is sharper on the surgical side than the stabbing side, which is true and completely beside the point if your concern is who holds the knife.

Why it matters: Two weeks ago, the Commerce Department pulled a live model offline. This week, the White House pre-screened a model's customer list before launch. The escalation is unmistakable. We are watching the emergence of a government approval regime for frontier AI capabilities, one that currently operates through informal requests rather than statutory authority but produces the same result: companies cannot release their strongest models to the general public without Washington's permission. Every major lab now knows that their next frontier release will require a conversation with the government before it requires a conversation with their sales team. That changes incentive structures throughout the industry, from how companies fund safety research (as compliance cost, not product differentiation) to how they price models (government-approved customers first, everyone else later, at a premium or a delay).

Why it might not: OpenAI's public pushback is notable. The company is not challenging the substance of the government's request but is explicitly arguing against its permanence, a negotiating position that suggests OpenAI expects broader release within weeks, not months. If Sol reaches general availability by mid-July with no incident, this episode may read more like a speed bump than a structural change. The administration also has no statutory framework for gating model releases, which means its leverage depends entirely on the willingness of companies to comply voluntarily. That willingness has limits, especially for a company approaching an IPO that needs to demonstrate broad commercial traction to public market investors.

2. Google Lost the Transformer and AlphaFold in the Same Week

Noam Shazeer announced on June 18 that he was leaving Google to join OpenAI as lead for architecture research. Shazeer co-authored the 2017 paper “Attention Is All You Need,” which introduced the Transformer architecture that underlies every major large language model in production today. He left Google in 2021 to co-found Character.AI, which Google brought back in 2024 through a $2.7 billion licensing deal, primarily to get Shazeer back. He served as vice president of engineering and technical co-lead for Gemini. OpenAI CEO Sam Altman posted that Shazeer was “one of the people I have most wanted to work with since the very beginning of OpenAI. Only took 10 years.”

Two days later, on June 20, John Jumper confirmed he was leaving Google DeepMind for Anthropic after nearly nine years. Jumper won the 2024 Nobel Prize in Chemistry alongside DeepMind CEO Demis Hassabis for creating AlphaFold, which has predicted the 3D structures of more than 200 million proteins and materially accelerated drug discovery worldwide. He ran DeepMind's AI for science division. Then Bloomberg reported that two additional senior researchers, Jonas Adler and Alexander Pritzel, were also departing for Anthropic. Four departures. One week.

Alphabet shares fell 5.1% on Monday, June 22, erasing roughly $240 billion in market capitalization, its worst trading day in over a year. The benchmarks were supposed to be the headline. Google had launched Gemini 2.5 Pro with Deep Think that same day, posting scores above every publicly available model: 82.4% on GPQA Diamond, 89.8% on MMLU-Pro, and a 2-million-token context window that can ingest entire codebases in a single session, a technical achievement so significant that in any other week it would have dominated the conversation, but in a week where the people who built the underlying architecture were walking out the door, the benchmarks read less like a triumph and more like a farewell gift.

Why it matters: Google paid $2.7 billion to bring Shazeer back in 2024. Twenty-two months later, he left for the company Google is trying to catch. That is not a talent management failure; it is a gravitational shift, one whose implications extend beyond any individual departure, because it demonstrates that the pull of the frontier labs now exceeds what even the world's most resource-rich technology company can counterbalance with compensation packages, research budgets, and the promise of deploying work at Google scale. Hassabis himself acknowledged the dynamic at an event in Cannes: “It’s the most ferociously competitive [talent market] it’s ever been in the tech industry.” Correct. But incomplete. Google is losing researchers to both OpenAI and Anthropic. Neither is losing researchers to Google. The talent flow is one-directional, and $240 billion in market value disappeared because investors noticed.

Why it might not: Google has the deepest research bench in AI and is the only company simultaneously operating frontier models, a global cloud platform, the dominant mobile operating system, and the default consumer search engine. Hassabis pointed out that Google wins “our fair share of the top talent.” Four departures from a research organization with thousands of scientists may not impair Gemini's trajectory, especially if the institutional knowledge embedded in Google's training infrastructure and data flywheel matters more than any individual contributor. Gemini 2.5 Pro Deep Think's benchmarks are real, and they launched the same week the researchers left, which means the models those researchers helped build are already in production. The question is not whether Google can ship competitive models today but whether it can continue to do so after the architects leave.

3. OpenAI Built a Chip. The Timing Is Not Coincidental.

On June 24, OpenAI and Broadcom unveiled Jalapeño, a custom inference processor designed from scratch around the specific requirements of running large language models in production. The chip is an ASIC, an application-specific integrated circuit, built for a single task: processing AI queries after models have already been trained. Unlike Nvidia's general-purpose GPUs, which handle training and inference across every workload type, Jalapeño is optimized for the memory movement patterns, serving loops, and latency requirements that define ChatGPT, Codex, and the API.

The development timeline is striking. Nine months from design to tapeout. That is roughly half the industry standard for advanced semiconductor design, a compression that OpenAI President Greg Brockman attributes to something recursive: the company's own AI models accelerated the process. “The degree to which our models have been able to accelerate it was very surprising to us,” he told CNBC. Engineering samples are already running production workloads, including GPT-5.3-Codex-Spark. Deployment at gigawatt scale is planned for late 2026, as part of a broader 10-gigawatt agreement with Broadcom through 2029. Broadcom expects AI semiconductor revenues to exceed $100 billion in 2027.

Why it matters: The official narrative frames Jalapeño as a cost-reduction play, and the economics are real. Inference is the most expensive line item in OpenAI's operating budget, and a purpose-built ASIC that delivers better performance per watt than general-purpose GPUs directly improves margins. But the timing tells a different story. OpenAI announced its first custom silicon two weeks after the government pulled Anthropic's models offline and in the same week Washington gatekept its own model release. When a government can revoke your model's availability overnight, owning your inference stack is not a cost optimization. It is a sovereignty play. The company that controls its own silicon, its own data centers, and its own model weights has one fewer point of external dependency that can be leveraged against it. That is the strategic context Broadcom investors should be pricing, not the inference cost savings.

Why it might not: Custom ASICs are a proven strategy for companies with predictable, high-volume workloads, but they carry architectural lock-in risk. If model architectures shift significantly (from Transformers to state-space models, for instance, or toward mixture-of-experts designs with different memory access patterns), a chip optimized for today's inference patterns may underperform tomorrow's requirements. Nvidia's advantage has always been generality: its GPUs handle whatever workload the industry throws at them. OpenAI is betting that it knows its own workload well enough to trade generality for efficiency. That bet is correct today. Whether it remains correct through 2029, the length of the Broadcom commitment, depends on whether model architecture stabilizes or keeps shifting.

4. Five Intelligence Agencies Said the Timeline Is Months

On June 22, the cybersecurity agencies of all five Five Eyes nations (the US, UK, Canada, Australia, and New Zealand) issued a joint statement warning that frontier AI models will “fundamentally transform both offensive and defensive cyber capabilities” within months, not years. The statement, signed by the heads of each country's national cybersecurity agency, including NSA and CISA leadership, urged organizations to treat cyber resilience as “a leadership priority, not a technical function.”

The agencies identified three specific weaknesses that AI will exploit first: weak identity controls, unpatched legacy systems, and unnecessary internet connectivity. Acting CISA Director Nick Andersen stated that “adversaries are already using AI to move faster and more effectively. Defenders must do the same.” Separately, the NSA's director testified that Anthropic's Mythos model had identified vulnerabilities in nearly all of the agency's own classified systems within hours during a controlled assessment.

In a related development, Representative Nathaniel Moran of Texas introduced the AI Incident Reporting Act on June 25, which would mandate that AI companies report dangerous capabilities, security breaches, and safety incidents to the Commerce Department within seven days, with penalties of up to $2 million per day for noncompliance.

Why it matters: Five Eyes joint statements are rare. Deliberate. The last comparable warning, in 2023, concerned Chinese state-sponsored intrusion campaigns. This one goes further by specifying a timeline (“months”) and by naming the defensive posture that organizations need to adopt before the window closes. Notice the convergence: in the same week, the intelligence community warned that AI-enabled cyberattacks are imminent, the White House gatekept a model release over cybersecurity concerns, and Congress introduced mandatory incident reporting legislation with $2 million per day penalties. Three independent branches of government reached the same conclusion through three different mechanisms within seven days of each other, a level of alignment on AI risk that has not occurred before and that suggests the internal assessments being briefed behind closed doors are substantially more alarming than anything in the public statement.

Why it might not: Intelligence agencies have institutional incentives to amplify threats, because threat inflation expands their budgets and their authority. The “months” timeline is vague enough to be unfalsifiable. If AI-enabled cyberattacks do not materially increase by year-end, the agencies can argue the warning prompted defensive action that prevented the predicted attacks, a circular logic that makes the prediction immune to disconfirmation. The NSA testimony about Mythos finding vulnerabilities in classified systems is alarming in isolation but requires context: sophisticated state actors have been finding those same vulnerabilities for decades using non-AI methods. The question is not whether AI can find vulnerabilities (it obviously can) but whether it lowers the cost of exploitation enough to change the threat landscape quantitatively. The Five Eyes statement asserts that it does. The evidence base for that assertion is not public.

5. SpaceX Bought the Code Editor. Now It Wants the Code Platform.

SpaceX finalized its $60 billion acquisition of Cursor, the AI coding startup, making it the most expensive software acquisition in history by a company that is still primarily known for rockets. The deal, which was first reported in April when the companies began sharing infrastructure, gives SpaceX access to Cursor's developer base, its training data, and its model development team. SpaceX had previously absorbed xAI, Elon Musk's AI lab, before its IPO, bringing frontier model development in-house.

On the day the acquisition closed, Cursor announced a new 1.5-trillion-parameter Composer model, trained from scratch on SpaceX's Colossus data center infrastructure, shipping in the coming weeks as part of both Cursor and Grok Build, SpaceX's own coding agent. Also announced: Origin, a code-hosting platform aimed squarely at Microsoft's GitHub. The vertical integration is explicit and complete. SpaceX now controls the data centers (Colossus I and II), the AI model (Grok), the coding tool (Cursor), and the code-hosting platform (Origin). Silicon to developer workflow. One company.

Why it matters: GitHub has over 100 million developers and sits at the center of modern software development. SpaceX is building a direct competitor backed by infrastructure that Cursor lacked as an independent company. The $60 billion price tag is eye-watering, but it buys something GitHub does not have: a company that controls its own training compute and its own frontier model, which means it can iterate on its coding AI faster than any platform that depends on third-party model providers. If Grok Build achieves Claude Code-level performance and ships natively inside a GitHub competitor, the developer tools market faces a structural disruption that was not on anyone's 2026 roadmap. The broader signal is that Musk is building the same kind of vertically integrated technology conglomerate that characterized the early decades of American industrialism, where the same company owned the factory, the railroad, and the steel mill.

Why it might not: Developer platforms have extreme network effects. GitHub's moat is not its features but its social graph: the pull requests, the issue trackers, the CI/CD integrations, the npm packages, the stars and forks that constitute a developer's professional identity. Origin starts from zero on all of these dimensions. Cursor itself was facing existential pressure before the acquisition, as both Claude Code and OpenAI's Codex were eroding its competitive position. SpaceX may have overpaid for a company whose independent viability was uncertain, betting that integration with its infrastructure would solve problems that were, in fact, product and distribution problems. The 1.5-trillion-parameter model is large, but parameter count stopped being a reliable proxy for capability two years ago.

6. Meta Made AI Glasses a Mass-Market Product

On June 23, Meta launched its first self-branded smart glasses, simply called Meta Glasses, at a starting price of $299. The glasses are built in partnership with EssilorLuxottica but do not carry the Ray-Ban or Oakley brand. They come in 26 style combinations across three frame shapes, including an oval model co-designed with Kylie Jenner. All are prescription-compatible. The glasses feature hands-free photo and video capture, open-ear audio, live translation, pedestrian navigation, and Meta AI powered by the company's new Muse Spark model, which replaces LLaMA as the on-device intelligence layer.

EssilorLuxottica disclosed that it sold more than 7 million AI glasses in 2025, up from 2 million combined in 2023 and 2024. At $299, the new Meta Glasses are $80 cheaper than the Ray-Ban Meta entry point ($379) and $1,896 cheaper than Snap's AR Spectacles ($2,195). Apple has reportedly pushed its own smart glasses competitor to late 2027. Google launched Intelligent Eyewear in May. Samsung is developing Galaxy Glasses with Google and Qualcomm for the second half of 2026.

Why it matters: The $299 price point is the line. Below it, smart glasses are an impulse purchase that competes with AirPods. Above it, they are a considered technology purchase that competes with tablets. Meta just crossed that line from the right direction, while every competitor except Google remains above it. Seven million units in 2025 on a $379+ ASP means the market exists. The question was always whether it could scale beyond early adopters, and pricing is the mechanism that unlocks mass adoption in consumer electronics, not features. The Kylie Jenner collaboration signals that Meta is positioning these as fashion accessories first and technology products second, which is exactly the framing that killed Google Glass in 2014 when Google tried it the other way around. Meta has the distribution (Instagram, Facebook), the AI model (Muse Spark), and now the price point. Those three together are a more formidable competitive position than any individual technical specification.

Why it might not: Seven million units is a rounding error relative to the 1.2 billion smartphones shipped annually. The use cases demonstrated so far, calorie estimation, language translation, museum recommendations, are incremental conveniences, not the kind of transformative applications that drove smartphone adoption. The Muse Spark model's capabilities at launch are comparable to what the existing Ray-Ban Meta glasses already do, which means the $299 price is subsidizing a form-factor bet, not a capability leap. If the AI experience does not improve meaningfully with future software updates, the glasses risk becoming another gadget that people try for a month and then leave in a drawer. The fashion play also cuts both ways: fashion products have seasonal cycles and brand fatigue that technology products can avoid, and Kylie Jenner's endorsement appeal is inversely correlated with the enterprise and professional demographics that typically drive sustained technology adoption.

7. The Memory Chip Company That Almost Died Now Rules South Korea

On June 22, SK Hynix surpassed Samsung Electronics to become South Korea's most valuable company for the first time, reaching a market capitalization of approximately $1.35 trillion. Samsung had held the top position since 2000. SK Hynix shares have risen more than 340% in 2026 alone, compared with Samsung's 200% gain. Both companies joined the $1 trillion club in May, alongside US rival Micron Technology. Memory chip prices doubled in Q1 2026 from the prior quarter and are forecast to climb up to 63% in Q2, driven by AI data center demand that has constrained supply for every other application that uses memory, from smartphones to automobiles.

The overtaking is the payoff from a 14-year bet. In 2012, SK Group acquired the struggling Hynix Semiconductor in a deal widely criticized as financially reckless. Rather than competing head-to-head with Samsung in conventional DRAM, SK Hynix invested heavily in high-bandwidth memory (HBM) chips, a niche product that could feed data to processors faster than anything else on the market but had limited customers. The company released the world's first HBM product with AMD in 2014, stumbled with the second generation, and nearly abandoned the technology in 2019 when demand collapsed. Executives debated halting HBM development entirely before deciding to double down, investing over 800 billion won in a packaging facility in Icheon.

Then Nvidia needed HBM for its AI accelerators, and the niche that nearly killed the company became the center of the most lucrative semiconductor market in history.

SK Hynix now holds a 57% share of the HBM market, its operating profit margin is expected to reach 70% in 2026, and its stock has risen more than 340% since January, a trajectory that makes it difficult to argue the AI infrastructure boom is a speculative bubble rather than a structural reallocation of global compute spending.

Why it matters: SK Hynix's ascent is the clearest demonstration that the AI hardware stack has a different set of winners than the conventional computing stack. For 25 years, Samsung dominated because it made the best commodity memory chips for smartphones and laptops. SK Hynix now dominates because it bet on the memory chips that AI accelerators specifically require. The distinction matters because it reveals where the structural demand is: not in consumer devices, where growth has plateaued, but in data center infrastructure, where every major technology company is spending at rates that make their own CFOs visibly uncomfortable. Memory, not processing power, is emerging as the binding constraint in AI infrastructure, a development that venture capital investors and public market analysts are only beginning to price into their models of the AI supply chain.

Why it might not: Market cap inversions driven by a single product cycle can reverse. Samsung is investing aggressively in its own HBM production and has the manufacturing scale to close the gap. If Samsung's next-generation HBM chips achieve yield parity with SK Hynix, the premium that investors are paying for SK Hynix's first-mover advantage could compress rapidly. The 70% operating margin is also a signal to every competitor, including upstarts, that HBM production is extraordinarily profitable, which will attract capital and capacity that eventually brings margins back toward industry norms. SK Hynix's bet paid off spectacularly, but the payoff is now attracting exactly the competitive response that payoffs always attract.

The Thing Nobody Is Talking About: ChatGPT Lost Its Majority the Same Week Washington Chose Its Successors

Sensor Tower's State of AI 2026 report, published in mid-June, contains a data point that landed with surprisingly little commentary: ChatGPT's share of the global AI assistant market fell below 50% for the first time. By the end of May, it stood at 46.4%. Gemini holds 27.7%. Claude holds 10.3%. The rest, Grok, Perplexity, DeepSeek, Meta AI, collectively account for less than 16%.

Place that number next to what happened this week and the dissonance is immediate: the AI assistant market is fragmenting while the government is consolidating control over which models can ship. Users are migrating between platforms based on capability, trust, and values alignment (Sensor Tower found that OpenAI's DoD deal triggered a measurable spike in ChatGPT uninstalls). The market is doing what markets do: distributing attention across competitors who earn it.

The government, simultaneously, is consolidating access. It is deciding which companies can release which models, to which customers, under which conditions. It is treating the most capable AI as requiring gatekeeping that resembles pharmaceutical approval or weapons export licensing.

Those two trends are headed for a collision. A fragmenting market produces more choices. A government-gated release regime produces fewer. Right now, the gatekeeping applies only to frontier capabilities, the models sitting at the cybersecurity threshold. But thresholds move. Right now, the gatekeeping applies only to frontier capabilities, the models sitting at the cybersecurity threshold. But thresholds move. Today's frontier is tomorrow's baseline. The government that gates GPT-5.6 Sol today will eventually face the question of whether to gate the open-weight models that approach Sol's capabilities next year, and the Chinese models that already benchmark within single digits of the American frontier. The tools for that gatekeeping do not exist yet, because you cannot gate-keep software that runs on someone else's hardware in a country that does not recognize your export controls. That asymmetry, a government that can restrict its own companies but not their foreign competitors, is the structural tension that no one in Washington has publicly addressed this week, even as both trends accelerated.

What You Can Do

If you lead an AI team: Model vendor risk is no longer theoretical. Anthropic's models went dark two weeks ago. OpenAI's latest model launched behind a government approval gate. Build inference routing across at least three providers and test failover monthly. The cost of redundancy is a fraction of the cost of your pipeline going dark because a model you depend on becomes the next subject of a government intervention.

If you invest in semiconductors: The Jalapeño announcement confirms that the shift from general-purpose GPU to custom inference ASIC is accelerating. Broadcom already has custom silicon deals with OpenAI, Anthropic, Meta, Google, and ByteDance. Nvidia's H100 and Blackwell remain dominant for training, but the inference market, where the recurring revenue lives, is diversifying faster than consensus models reflect. If your portfolio thesis depends on Nvidia maintaining inference-market share, stress-test it against a scenario where every major lab ships its own silicon by 2028.

If you are a developer choosing tools: The SpaceX-Cursor deal changes the competitive landscape for AI coding assistants. Before this week, Cursor was an independent tool that worked with every model provider. It will now become a distribution vehicle for Grok and a competitor to GitHub. If you currently depend on Cursor, evaluate whether you are comfortable with SpaceX controlling your development environment. If you are evaluating alternatives, Claude Code and OpenAI Codex both offer comparable capabilities without the platform lock-in risk that a SpaceX-owned tool will carry once Origin launches.

If you run enterprise security: The Five Eyes warning is not hypothetical guidance. It is a formal assessment from five national intelligence agencies that the threat landscape will change materially within months. The three defensive priorities they named (identity controls, patching, reducing internet exposure) are not novel, but the urgency is. If your organization has a quarterly patching cycle, compress it to monthly. If you have systems exposed to the internet that do not need to be, take them offline now, before the window the agencies described closes.

Limitations

This analysis relies on public reporting, company statements, and market data. The White House's request to OpenAI has not been published as a formal directive; the arrangement appears to operate through informal coordination rather than statutory authority. OpenAI's characterization of Sol's cyber capabilities relative to GPT-5.5 is self-reported and has not been independently verified. The $240 billion figure for Alphabet's market cap decline is based on intraday trading on June 22 and may differ from close-to-close measurements. SK Hynix's market capitalization briefly surpassed Samsung's on an intraday basis and the relative positions may fluctuate. The Five Eyes timeline of “months” is not further specified. The NSA testimony regarding Mythos finding vulnerabilities in classified systems is cited through secondary reporting and the primary testimony is not public. SpaceX's $60 billion acquisition price for Cursor has not been confirmed through SEC filings. Sensor Tower's market share figures measure usage across desktop, mobile apps, and mobile web through a “true audience” metric whose methodology has not been fully disclosed. ChatGPT's sub-50% market share was reported in mid-June using May data. Notable stories not covered in depth: Google's Gemini 2.5 Pro Deep Think benchmarks, the AI Incident Reporting Act's full provisions, Google's FARO regulatory proposal, China's Z.ai GLM-5.2 model reaching $128 billion market cap, and Dario Amodei's call for binding AI regulation.

This is the ninth installment of “The Biggest Things in AI This Week.” Previous editions: June 15 · June 8 · June 1 · May 24 · May 17 · May 11 · May 4 · April 27