The Washington List: AI by Decree — Silicon Valley Confidential

MY TAKE

Last week I wrote that 2026 was the year frontier AI got nationalized de facto without anyone signing a decree. This week we’ve seen that Washington doesn’t just ration; it now modulates.

On July 1, the export controls on Anthropic’s Fable 5 and Mythos 5 were lifted, and the company started redeploying them on the 2nd, while OpenAI’s GPT-5.6 stayed locked — but Google was given the green light to launch Gemini 3.5 Pro in July. Trump has gone from flipping the switch on and off to deciding whose switch gets flipped, and when.

The way I see it, the important issue isn’t the restriction, but the price of getting out of it. Anthropic — I have no proof but no doubts either — got its models back by signing conditions.

There’s talk of “proactively detecting and addressing security risks,” of collaborating on the standards for its next models, and of reporting “malicious activity” to the government.

I don’t know, Rick — I love frontier models, but we need a plan B.

They’ve pulled Claude Sonnet 5 out of the hat at $2/$10 per million tokens — they say it’s almost an Opus, at bargain-bin pricing, so you can run agents all day. I’ve tried it raw and I don’t like it, but it works well in PaellaDoc — then again, in PaellaDoc even Kimi 2.7 works well.

Together AI has raised $800 million to sell inference on open models below the closed labs’ prices. This is gathering momentum; for Europeans worried about DPAs and GDPR, I recommend looking at Cortecs.ai

I believe the open layer of efficiency and commodity — neoclouds, open weights and sovereign money that exists precisely to depend neither on the Washington list nor on the labs’ rate card — is heading up.

My read for the start of the second half: the enterprise buyer’s question is no longer “which model is smarter.”

It’s “how much of my stack do I want to expose to US discretion.” This spring’s market shift was toward free models like GLM 5.2.

I think the money poured into OpenAI and Anthropic is the risk that, sooner or later, a public market will have to put to the test.

Happy July — I’ll leave you with my minions if you want more detail.

The day Washington unlocked one lab and left the other in the cage

Last edition’s cliffhanger resolved this week, and it did so in the most revealing way possible: not with a blanket “yes” or “no,” but with different decisions for each company. The state-license regime that debuted seven days ago proved to be discretionary and personalized. Timeline.

June 26 (starting context): OpenAI unveils the GPT-5.6 family — Sol, Terra, Luna — and limits it to some twenty “trusted partners” at the government’s request, public protest included. The same day, Commerce releases Anthropic’s Mythos 5 to more than 100 US institutions, but leaves Fable 5 — its most powerful public model — still locked.

June 29: signals of the endgame start leaking. It’s reported that Fable 5 is about to return, that Gemini 3.5 Pro has been cleared for a July launch, and that GPT-5.6 remains padlocked. The asymmetry begins to take shape.

July 1 (late Tuesday) — the key moment: Commerce Secretary Howard Lutnick notifies Anthropic that he has withdrawn the export controls on Fable 5 and Mythos 5. It no longer needs an export license — with conditions: Anthropic must proactively detect and address security risks, work with the government on the standards for its future models, and report “malicious activity.” The original trigger for the block, per the reconstruction, had been fear of a security vulnerability in Fable 5.

July 2: Anthropic starts restoring global access to both models. It doesn’t return as-is: it accompanies the redeployment with deeper cyber safeguards, a draft “AI jailbreak severity framework,” and a HackerOne program for researchers to report potential cyber jailbreaks. The company thanks users “for their patience.” OpenAI’s GPT-5.6 remains restricted.

Why it matters. A week ago the headline was “the state rations the frontier.” This week the headline is finer and more unsettling: the state chooses lab by lab. Anthropic got out before OpenAI not for being worse or more harmless — Mythos is its most powerful cyber model — but because it negotiated and signed a set of conditions tying it to the national security machinery. OpenAI, which protested publicly against the regime, still has its flagship grounded. Google, which made no noise, got its takeoff clearance.

The signal: in the era of state-licensed AI, each lab’s individual relationship with the Commerce Department is a competitive asset as decisive as its best benchmark. Availability is no longer a product variable, or even a general permission variable: it’s a bilateral negotiation variable. For the enterprise architect, this week’s lesson tops last week’s. It’s not enough to ask whether your vendor is “on the list”: you have to ask what it signed to get in, what surveillance obligations it took on over your usage, and what happens when the next model enters the same funnel. The one that redeployed first did so by agreeing to be, in part, the state’s eyes and ears inside its own product.

POWER MOVES

Google clears Gemini 3.5 Pro for July and launches it straight into its own brain drain

The move isn’t a hire: it’s the decision to ship product in the middle of a hemorrhage. On June 29 it emerged that Commerce had authorized the launch of Gemini 3.5 Pro in July, just as the Google DeepMind talent exodus — the big power move of last edition, with Alphabet shedding about $270 billion in market cap over two sessions — kept widening. In the same period the departures of Noam Shazeer (transformer co-author, to OpenAI), John Jumper (Chemistry Nobel for AlphaFold, to Anthropic) and key Gemini contributors like Jonas Adler and Alexander Pritzel were confirmed or consolidated; the running count is around six researchers in five months, and Denny Zhou, founder of Google Brain’s reasoning team, has been at Meta for months. To top it off, Google’s current models — Gemini 3.1 Pro, 3.5 Flash — frequently show up outside the leaderboard top-5.

Google’s play is ship-to-reassure: send the model to defend the narrative rather than pause to stabilize the bench. Demis Hassabis insists Google is “still winning the talent war”; the market and the rankings say otherwise. The signal: in 2026, launch cadence has become a hostage-signal to the market. With OpenAI and Anthropic shipping models and features almost weekly, Google can’t afford a visible gap at the frontier while its research bleeds out. It launches because not launching would confess the problem. But shipping a model doesn’t replace a Nobel laureate, and that’s exactly the moat being arbitraged.

Aramco Ventures leads Together AI’s $800 million: the Gulf buys the open-source hedge

On July 1, Together AI — a “neocloud” that rents Nvidia GPU clusters and serves open models (DeepSeek, MiniMax, Kimi) below the price of OpenAI’s and Anthropic’s closed systems — closed an $800 million Series C at an $8.3 billion valuation, led by Aramco Ventures, with Nvidia, Vista Equity, General Catalyst and others. That’s a leap from $3.3 billion 16 months ago. The company says it has more than $1.15 billion in annual bookings and plans to multiply its capacity ~50x in five years.

The signal: when you can’t win the model race and don’t want Washington rationing you, you buy the picks and shovels. Together AI is doubly antifragile against the week’s two threats: it arbitrages the closed labs’ price downward and, by operating on open weights, completely sidesteps Commerce’s licensing funnel — open models don’t need lab-by-lab export permission. That the investment arm of the world’s largest oil company is writing the lead check on that bet isn’t anecdotal: it’s sovereign capital positioning itself to own the physical and commercial substrate of the AI others think is only played in the model weights.

MONEY TALKS

The first half closes at $510 billion — an all-time record — and 43% went to two companies

The macro number of the week landed July 2: global venture funding reached $510 billion in the first half of 2026, more than was invested in all of 2025 ($440 billion) and far above the previous semiannual record (about $375 billion in H2 2021). The breakdown is brutal: Q1 was $305 billion (the biggest quarter in history) and Q2 was $205 billion (the second biggest). OpenAI and Anthropic together took $217 billion, 43% of all semiannual capital; Anthropic’s $65 billion Q2 round alone ate close to a third of the quarter’s global capital. On exits: 32 companies went public above $1 billion and there were 24 M&A deals above that mark, for $113 billion combined. Sixteen companies raised megarounds of $1 billion or more in Q2 alone.

The report itself warns: an “unprecedented share” of the half’s capital flowed to just two companies. The signal: the AI boom has become a two-name index. This isn’t a broad euphoria lifting all boats, but an asymmetric barbell where OpenAI and Anthropic carry the entire market’s AI beta. The systemic risk isn’t a bubble popping everywhere: it’s that the perceived health of a $510 billion sector depends on two private companies — both headed for public scrutiny — continuing to validate their valuations with real growth and not just the next round.

The IPO window flew open on July 1: Bending Spoons and Lime list, pop, then wobble

The exit market for non-AI tech reopened this week, and the result is a lesson in nuance. Bending Spoons — the Italian firm that buys and operates digital businesses: AOL, Vimeo, Evernote, WeTransfer, Brightcove, Eventbrite — priced its IPO at $29 per share, above the $26-28 range, raising about $1.68 billion. It debuted July 1 with a 40% pop to $40.50… and fell 8.5% the next day to about $37. Lime — the Uber-backed e-scooters and e-bikes — priced at $25 (top of its range), opened July 1 at $27 (+8%), at a $1.8 billion valuation and $182 million raised, with Uber coming in as anchor investor for up to $20 million. The detail an underwriter shouldn’t miss: Lime acknowledges $845.8 million of debt coming due within twelve months and admits that, without the IPO, it “could go bankrupt.”

The signal: the window is open, but the public market is reading it as “grab liquidity while you can,” not as conviction. The day-two fades (Bending Spoons) and Lime’s going-concern language say the public investor is selective and distrustful. The contrast with the $510 billion in private money is the real capital story of 2026: private appetite for AI is indiscriminate; public appetite for the rest of tech is surgical. Two thermometers, two temperatures.

PRODUCT SECRETS

Claude Sonnet 5: Anthropic makes the agent cheap enough to leave running all day

On June 30 Anthropic launched Claude Sonnet 5, and the substance isn’t the intelligence: it’s the economics. The model offers performance close to Opus 4.8 — the flagship — but is the most agentic Sonnet to date (it plans, uses a browser and terminal, operates autonomously) at an introductory price of $2 per million input tokens and $10 output through August 31 (then $3/$15). It becomes the default model on Free and Pro plans. The target is clear: this isn’t an assault on the capability ceiling, it’s a price collapse for “good-enough autonomy” to capture the recurring budget of running agents, which is where the month-after-month spend lives. It fits the shift CNBC documented in late June: users moving from “tokenmaxxing” — hoarding capacity — to squeezing efficiency. The signal: the battlefield has moved from the leaderboard to cost per task. The lab that solves the same job with less output wins the budget, even if it ties — or narrowly loses — on the benchmark. Anthropic just put commodity pricing on its second-best model so the CFO has no excuse.

Claude Science: Anthropic plants the flag on the beach where its Nobel laureate just landed

The same June 30, Anthropic unveiled Claude Science, a beta workbench for scientists: it unifies research tools and packages, generates auditable artifacts, enables reproducible analyses and offers flexible compute access on macOS, Linux and remote environments. It’s positioned head-on against OpenAI’s GPT-Rosalind and Google’s Isomorphic Labs. The timing is no accident: it arrives days after John Jumper — Nobel laureate for AlphaFold — landed at Anthropic from DeepMind. The signal: the labs are racing to own the scientific workflow because it’s the most valuable, stickiest, most defensible vertical — and, incidentally, the one with the best narrative (“AI does science”). Claude Science is the flag planted exactly on the beach Jumper just stepped onto. In a talent war where signing the star costs hundreds of millions, turning their arrival into a sellable product the same week is making the hire pay rent from day one.

Google cheapens images and AWS sets itself up as the neutral switchboard: the application layer commoditizes

While the flagships hog headlines, application infrastructure kept commoditizing. Google launched two image generation models on June 30 — Gemini 3.1 Flash Image and Gemini 3 Pro Image — pushing image pricing toward the floor (from $0.50 per million input tokens on the Flash). And AWS consolidated Claude, Amazon Nova, Meta’s Llama, Mistral, AI21 and xAI’s Grok on Bedrock under a single contract and single billing. The signal: Amazon is positioning itself as AI’s neutral switchboard — the Switzerland that tolls everyone without marrying anyone. And that neutrality is precisely the customer’s best defense against this week’s Bombshell: if your model vendor can be rationed by Washington overnight, being able to switch labs from a single contract stops being a convenience and becomes a business continuity strategy.

REAL NUMBERS

AI-driven layoffs lead for the fourth straight month, and tech is already 31% of the half. The Challenger report released July 3 delivers the year’s uncomfortable stat: AI was the most-cited reason for US job cuts for the fourth month running, something without precedent. The tech sector eliminated about 139,156 positions, up 83% year over year, and accounted for 31% of all first-half layoffs. The 2026 running total is around 185,894 workers across 267 events — close to 1,000 per working day — with 56% of events citing AI or automation. The implication: the layoff-capex paradox is no longer an episode, it’s a four-month trend. The same companies cutting — Amazon, Microsoft, Alphabet, Meta — are guiding toward some $700 billion of AI capex in 2026. The productivity return that would justify both at once still isn’t showing up in the margins, and that gap between story and income statement is what public scrutiny — with OpenAI and Anthropic headed to the exchange — will sooner or later turn into a board question.

Tesla: 480,126 Q2 deliveries, +25% year over year, beating consensus by ~74,000 units. On July 2 Tesla reported 480,126 vehicles delivered (467,762 Model 3/Y and 12,364 other models), topping the ~406,000 estimate by around 18% and marking its first quarter of annual growth since the 2023 peak. It also deployed 13.5 GWh of energy storage. Full financial results arrive July 22. The implication: the “EV demand is dead” narrative takes a hit, but read the fine print: the comparison bar was low (the first growth in years) and the real acid test isn’t volume, it’s the margin and average price we’ll see on the 22nd. A delivery beat brightens the headline; the margin will say whether it was bought by cutting price.

Nvidia and China: 75,000 chips per customer authorized… and zero delivered. Per this week’s reconstruction (The Wire China, July 2), Commerce authorized about ten Chinese firms — Alibaba, Tencent, ByteDance — to buy Nvidia’s H200 with a cap of 75,000 units per customer, but not a single chip has shipped: the operation remains in legal limbo over US-China tech rivalry and Beijing’s new supply chain rules, while Huawei gains share in the local market. The implication: China’s “reopening” for Nvidia exists on paper, not in the cash register. The real revenue remains hostage to two governments, and every quarter of delay is share Huawei consolidates at home. The same dynamic as the Bombshell — state permission as a business variable — operates here in hardware.

THE DRAMA

Anthropic exits the cage by signing; OpenAI stays inside after protesting: cooperation pays better than complaining

The Bombshell’s unlock left a competitive subtext that deserves its own spotlight. Anthropic got Fable 5 and Mythos 5 back on July 1-2 because it signed: it agreed to proactively detect security risks, co-write the standards for its next models and report “malicious activity” to the government. OpenAI, which in late June protested publicly against the regime — saying in writing that this government access process should not become the default standard — still has GPT-5.6 restricted this week. Two opposite strategies before the same referee: quiet cooperation versus public complaint. So far, cooperation gets paid first.

Estimated impact and the signal: in the state-license era, each lab’s relationship with the Commerce Department has become a competitive weapon. Anthropic turned a regulatory crisis into a two-week positioning advantage over its direct rival, at the price of accepting obligations that pull it closer to the national security machinery. The power dynamic is the one defining 2026: the state doesn’t just decide who sells — it rewards the best cooperator with access. That reorders the incentives of the entire industry — protesting is expensive, cooperating pays — with an uncomfortable implication for anyone hoping the labs would act as a counterweight to state power instead of competing for its favor.

Venice hits $1 billion selling “surveillance-free AI”: the vote of no confidence, unicorn edition

In parallel, a two-year-old Wyoming startup, Venice — which offers private, “surveillance-free” access to a range of AI models — closed a $65 million Series A led by Dragonfly, valuing it at $1 billion. The juicy detail is the timing: it reaches unicorn status the very week Washington tightens access to the frontier, and on the heels of last week’s scandal, when it emerged that Meta was training AI by surveilling its own employees. The signal: “privacy from the labs and from the state” is now a fundable thesis with a billion-dollar sticker. The underlying drama is the market’s implicit verdict: there are enough buyers who distrust the incumbents’ data practices for anti-surveillance to be a venture-scale wedge. In a week when power — state and corporate — concentrated over data, someone just put a price on the opposite.

THE WEEK AHEAD

Monday, July 6 — ICML 2026 kicks off in Seoul. The big machine learning conference, doubly relevant in a talent war: a source of research announcements and a hiring thermometer. Watch who presents under which affiliation — every logo change on a poster is a market signal.
The GPT-5.6 endgame. Anthropic exited the funnel by signing conditions; the open question is whether Commerce will give OpenAI the same treatment, and when. If GPT-5.6 unlocks with a similar term sheet, that confirms there’s a procedure; if it takes longer, that confirms there’s discretion. The asymmetry is the story.
Gemini 3.5 Pro, general availability in July. Launch authorized; date pending. Every day Google doesn’t ship its frontier, with OpenAI and Anthropic pushing and its bench walking out, is a narrative problem.
Remedies decision in the Google ad-tech antitrust case (no fixed date). Judge Brinkema has yet to hand down the remedy after ruling Google illegally monopolized ad-tech. The DOJ wants structural divestiture of AdX and open-sourcing the auction logic; Google wants behavioral remedies. It could drop any day and would be the most consequential Big Tech breakup in decades. Maximum-weight wildcard.
Wednesday, July 22 — Tesla Q2 financial results. The delivery beat is in; on the 22nd comes the real test: automotive margin and average price. That’s where we’ll see whether the growth was bought by cutting price.
Post-IPO trajectory of Bending Spoons (BSP) and Lime (LIME). Bending Spoons’ day-two fade and Lime’s going-concern language set the tone. If they hold, the non-AI exit window stays open for the second half; if they deflate, it closes fast.
US-China: does any H200 actually ship? Up to 75,000 chips per customer authorized, zero delivered. Any real shipment — or any new obstacle from Beijing — moves Nvidia’s revenue narrative and semiconductor sector sentiment.

REFERENCES

The bombshell — the asymmetric unlock:

Power moves:

Money — the semiannual record and IPOs:

Product:

Real numbers:

Drama:

The week ahead:

Silicon Valley Confidential is published weekly. Executive intelligence verified against primary sources. Compiled: July 5, 2026 | Period: June 28 – July 5, 2026