Last Week Ignite: June 28 to July 5, 2026
The Off Switch and the Glut
Happy Fourth of July, everyone. Two hundred fifty years ago, a group of people bet everything on an unproven idea: that ordinary citizens could govern themselves, and that liberty and justice were not gifts from a crown but rights worth declaring. The experiment is still running. That it has lasted a quarter millennium is remarkable. That it still demands something of us is the point.
On June 30, the US Commerce Secretary sent a letter to Tom Brown, one of Anthropic’s cofounders, and turned the company’s most powerful AI model back on. Eighteen days earlier the same government had ordered it off. The trigger had been a jailbreak, discovered by Amazon, that coaxed the model into writing working exploit code. The condition for restoration was a set of promises: Anthropic would hunt for security risks proactively, help write the standards its future models would be judged against, and report malicious use to Washington.
Read that sequence again slowly, because nothing like it has happened before. A frontier AI model, the commercial crown jewel of one of the most valuable private companies in the world, was switched off by government order and switched back on by government letter, all inside three weeks. The same week, OpenAI’s newest model stayed locked to roughly twenty government-approved organizations, unavailable in ChatGPT, waiting for its own letter.
For two years the AI trade has run on two assumptions. The first is that compute is scarce and will stay scarce, so anyone holding GPUs holds pricing power. The second is that frontier models ship when their makers decide to ship them. Both assumptions cracked this week, and the second one cracked in public, with a return address.
Who decides what ships
Start with the permission story, because it is the stranger of the two.
Anthropic’s Claude Fable 5 went back on sale globally on July 1. Its restricted sibling, Mythos 5, was restored to about a hundred US critical-infrastructure organizations. Anthropic says a retrained safety classifier now blocks the exploit that caused the freeze more than 99 percent of the time, a company-stated figure, and it opened a public bug bounty to find the next one. Alex Wissner-Gross, the Harvard physicist whose Innermost Loop newsletter tracks frontier AI, opened his July 1 post with the line “The Singularity just cleared customs,” and quoted security researcher Alex Stamos calling the whole episode a “huge own goal for the US.”
Meanwhile GPT-5.6, which OpenAI previewed on June 26 in three pricing tiers, spent the entire week gated. Same government, same month, opposite outcomes for the two leading labs.
Here is why this matters more than the news cycle suggests. Until June, the question “can I build my product on this model” was a commercial question. You checked the price, the rate limits, the terms of service. Now it is also a regulatory question, and regulators have demonstrated they will answer it differently for different companies in the same month. Experienced investors will tell you that the moment a single supplier can be shut off by forces outside your contract, dependence on that supplier stops being a convenience and starts being a disclosed risk. Every startup built as a thin layer on one frontier model just inherited a risk factor it cannot negotiate away.
The market is treating the Anthropic freeze as a one-off. The more likely reading is precedent given how strong models are getting now. A June executive order already sketches a voluntary pre-release review for frontier models. The thing to watch over the next month is whether voluntary hardens into mandatory, because if it does, the release calendar for the most important technology of the decade runs partly through Washington.
The glut nobody priced
The second crack was quieter in its mechanics and louder in the tape.
Around July 1, Bloomberg reported and CNBC confirmed that Meta will start selling its excess AI compute to outside customers, turning the largest private GPU hoard in the world into merchant supply. Meta’s stock rose 8.8 percent that session on roughly triple its normal volume. CoreWeave, a company whose entire business is renting GPUs, fell 10.8 percent. Nebius, same business, fell 12.4 percent. Micron, which makes the memory chips that feed those GPUs, fell more than 10 percent. Gil Luria at D.A. Davidson put the problem plainly: the neoclouds “rely on Meta for their growth and Meta may not need them anymore.”
One announcement, no product, no pricing page, and tens of billions of dollars in market value moved. That tells you how much of the AI trade was resting on the assumption that compute stays scarce forever. Meta is the second giant to break ranks; SpaceX had already begun renting compute to Anthropic and Google. When the biggest hoarders become sellers, scarcity is a policy choice, and policies change.
The confusing part is that physical scarcity is still real. On Micron’s earnings call in late June, CEO Sanjay Mehrotra said “sustained and strong industry demand, along with supply constraints, are contributing to tight market conditions and we expect these conditions to persist beyond calendar 2026,” and disclosed that every unit of high-bandwidth memory Micron will make in 2026 is already committed under contract. Memory and power are genuinely tight. What cracked this week is the premise that GPU rental specifically stays a seller’s market. Those are different bottlenecks, and the market spent Tuesday learning the difference.
Agents got cheap, with an asterisk
While the permission and supply stories played out, the price of intelligence itself dropped.
Anthropic released Claude Sonnet 5 on June 30, its most capable mid-tier model, at introductory pricing of two dollars per million input tokens and ten per million output, rising to three and fifteen after August. Its top model, Opus 4.8, costs five and twenty-five. On Anthropic’s own numbers, Sonnet 5 beats Opus on a terminal-use benchmark and on a knowledge-work benchmark while trailing it on hard software engineering. Every one of those figures is vendor-claimed until someone outside Anthropic reproduces it, and vendor benchmark leads rotated multiple times this week alone, so hold them loosely.
The asterisk sits in the fine print. Sonnet 5 uses a new tokenizer, the component that chops text into the units the model bills you for, and it maps the same input into 1.0 to 1.35 times more tokens than before. Cheaper per token can still mean pricier per task. Any founder recomputing their unit economics this week needs to run the math on tokens consumed per job, and the price per token second.
There is a deeper pattern underneath the pricing news. Two research claims surfaced this week, both secondhand and unverified, both pointing the same direction. ByteDance researchers reportedly found agent learning speed doubling every three months on a benchmark of theirs, and the UK AI Safety Institute reportedly showed a model’s capacity for long cyber tasks stretching from two hours to fourteen as its token budget rose twentyfold. If either result holds up, agent capability is partly a function of how much inference you can afford, the way a basketball team’s ceiling is partly a function of how many minutes its best player can stay on the floor. Cheaper inference literally buys smarter agents. Whoever controls cheap inference controls the rate at which agents improve.
Where the money went
Which brings us to the week’s largest venture round, because it is the same story told in capital.
Together AI, a platform that makes open-source models cheap to run, raised 800 million dollars at an 8.3 billion valuation, led by Aramco Ventures with Nvidia, Vista, General Catalyst, and Salesforce’s venture arm participating. The company disclosed annual bookings crossing 1.15 billion, with open-model usage tripling in a year. Its customer list is the tell: Cursor, Cognition, Decagon, the very AI application companies whose margins get eaten when frontier-model prices stay high. Decagon says its inference bill fell sixfold after moving over. When a Saudi state energy fund and the chipmaker itself co-underwrite a company whose product is lower inference prices, the smart money is betting the closed-model premium shrinks.
The second notable round was Chamath Palihapitiya’s 8090, an enterprise AI coding company, raising 135 million led by Salesforce Ventures, with Palihapitiya taking the CEO seat, his first operating job since Facebook. The company’s claims about translating eighteen million lines of COBOL for insurers are its own and unaudited. The signal is in what got funded: governed execution in regulated industries, where the audit trail is the product, rather than another general coding assistant. Note that Salesforce now anchors both 8090 and part of Together’s cap table. An incumbent is buying position in the layer where enterprise software gets built, which sharpens the acquire-or-compete question for every startup in that lane.
Zoom out to the whole week’s funding table and the pattern gets blunt. The largest round was Joulent at 1.75 billion for AI-oriented energy infrastructure, backed by National Grid’s venture arm. Then Together. Then a compliance company, then 8090. Power, compute plumbing, and compliance topped the table. Application software did not. A reminder worth repeating: funding is a price signal, and prices can be wrong, but the direction of the pricing is information. The market is paying up for the picks and shovels and grading the apps harder.
Down at the stage where seed investors live, five rounds worth naming printed inside the week: a 30 million seed for Nebex in space-economy settlement led by GV, a 19.5 million Series A for Pie in small-business customer acquisition led by Lightspeed, a 31 million Series A for Omen AI in data-center sensor monitoring, a 12.6 million seed for Queue in autonomous pharmacy robotics, and 7.5 million for Civ Robotics in outdoor surveying robots, the latter two led by AlleyCorp. Three of the five are physical AI with a clear commercial buyer. Seed pricing has not repriced downward in robotics; if anything the deployment-constrained physical companies are getting funded faster than software.
One small founder-behavior anomaly worth filing away. Omnea, a London AI procurement company, announced it will hand employees 250,000 dollars to openly plan their next startup instead of moonlighting in secret. One company, small sample. But it names something real about this talent market: the best operators now expect to found, and employers are starting to price retention against the seed market itself.
The macro that will not help you
The Bureau of Labor Statistics reported July 2 that June nonfarm payrolls rose 57,000 against a consensus of 115,000. Participation fell to its lowest level since March 2021, household employment dropped by half a million, and prior months were revised down. Wages still rose 3.5 percent year over year. This is a labor market cooling through weak hiring rather than layoffs, which is the awkward kind of soft: weak enough to worry about, firm enough on wages that the Federal Reserve feels no urgency.
And the Fed said as much. Chair Kevin Warsh, speaking at the ECB’s Sintra forum on July 1, said “prices are too high” and “we’re going to deliver price stability,” and declined to give any guidance on the July meeting. Rates sit at 3.50 to 3.75 percent and nearly half the committee still leans toward a hike this year. The second-half rate-cut thesis that a lot of 2026 growth-stage burn plans were quietly built on is dead for now. Worse for planners, the Fed has stopped pre-announcing its path, so repricing will arrive without warning when it arrives.
One mechanical event rounds out the week. SpaceX joins the Nasdaq-100 before the open on July 7, a fast-track inclusion just fifteen trading days after its June IPO, with J.P. Morgan estimating around 4.3 billion dollars of forced passive buying concentrated after the July 6 close. Index funds have to buy regardless of price. Whatever SpaceX shares do around those two days reflects flow mechanics, and tells you nothing new about the business.
For founders
Every founder building on AI should answer four questions this week, and the honest answers sort companies fast.
If your best model were switched off by government order tomorrow, what is your fallback, and how many days does it take to deploy?
What do your unit economics look like at Sonnet 5 pricing, after the tokenizer change, versus ninety days ago?
Do you own a workflow end to end, or are you a feature a platform can absorb the next time it ships enterprise controls?
If inference prices fall 30 percent over the next year, does that expand your margin or commoditize your product?
The winners from this week’s shifts are visible. Agent products whose economics were underwater at last quarter’s model prices just became viable. Routing and fallback infrastructure, the switching layer that lets a company move between models when one gets expensive or gets banned, went from nice-to-have to board-level topic in eighteen days. Tooling that watches, permissions, and rolls back autonomous agents gets more urgent every time agents get cheaper and longer-running, and this week they got both. And vertical AI in regulated industries, where the compliance trail is the moat, just watched 8090 validate the wedge with strategic capital.
The losers are equally visible. Undifferentiated GPU rental now competes with Meta’s cast-offs. Thin wrappers on a single frontier model carry policy risk no term sheet can hedge. Burn-heavy growth plans premised on cheap money in the second half need new premises. And startups selling cost dashboards around Claude just watched Anthropic ship those controls itself, which is the oldest platform story there is.
For LPs
The variables that governed this cycle are rotating. For two years the questions were who has capital and who has compute. The questions now are who has permission and what inference costs. Neither shows up in a quarterly mark until it suddenly does.
Three practical implications. First, early-stage exposure to capital-efficient companies that own their workflow ages better in a world where model access can be revoked and model prices can collapse; both happened, in both directions, in one week. Second, pacing beats speed while the Fed refuses guidance, because macro repricing will arrive unannounced and vintages that deployed patiently through 2026 will look smarter than vintages that sprinted. Third, late-stage private marks now carry public-market machinery inside them. SpaceX exposure is partly an index-flow instrument as of July 7. Anthropic’s valuation now embeds a demonstrated government off switch. Underwrite the risk factor, and treat the headline number as the output rather than the input.
For the venture business generally
The uncomfortable summary of the week is that two premises underwriting most AI theses since 2024 broke within four days of each other. Permanent compute scarcity broke on Tuesday when Meta became a seller. Unconditional model access broke on Monday when a government letter, and only a government letter, put Fable 5 back on the market. Any thesis document that contains the phrase “GPU scarcity” or assumes frontier models ship on lab timelines is due for a rewrite.
Value is migrating to three places: whoever is allowed to ship, whoever makes inference cheap, and whoever owns the workflow the model plugs into. The middle of the stack, the layer that merely resells intelligence made by others, got squeezed from above by policy and from below by price in the same week. That is the story to underwrite, and next week’s job is to see whether Washington makes it a pattern.
Sources for this issue include the Anthropic and Together AI announcements, the Commerce Department letter as reported by CNBC and Al Jazeera, BLS Employment Situation (July 2), Micron’s fiscal Q3 call, Nasdaq inclusion notices, Crunchbase funding data, and The Innermost Loop (July 1 to 4). Benchmark figures are vendor-claimed unless noted. Two capability-scaling claims (ByteDance EdgeBench, UK AISI cyber task horizons) are secondhand and unreproduced; they are flagged as such where cited.