SpaceX Paid $60B for Cursor and the Government Switched Off Anthropic's Best Model

MY TAKE

Folks, the frontier model business is being compressed from below and restricted from above, and it’s being financed on the public markets before the numbers add up. All three at once. Let’s take it piece by piece.

From below: China’s Z.ai has released GLM-5.2 with open weights and an MIT license, and it’s landed within one benchmark point of Claude Opus 4.8 on coding, at one sixth of the cost. When the gap between the best open model and the best closed one is a single point instead of a full generation, that’s not a technical curiosity: it’s margin erosion exactly where Anthropic and OpenAI monetize the most, agentic software development.

I put GLM through its paces this weekend and… I still see differences.. I’d call it a Sonnet 4.6, but it’s not bad.

In the US, Trump ordered Anthropic to suspend access to Fable 5 and Mythos 5 invoking export control authorities, and this week, six days later, the models were still switched off while the company negotiated their return with the Commerce Department. It’s the first time that lever has been used against a specific frontier model. The availability of a SOTA model has stopped being a product variable and become a geopolitical one. Any lab that ships its best work now carries that risk.

And that leaves us with… the financing: OpenAI and Anthropic racing toward an IPO at valuations approaching a trillion dollars while losing money hand over fist. If the readings of the confidential financials are correct, OpenAI was losing $1.22 for every dollar of revenue. They must know something we don’t.

Robinhood laid off 10% of its workforce and, for the first time in months, its CEO was extremely careful not to mention AI. We’ve spent a year watching CEOs attribute every cut to “artificial intelligence efficiencies” because that narrative traded up on Wall Street. When they start avoiding it, it means it has stopped trading up.

Laying people off while invoking AI at a business that’s growing has gone from a signal of efficiency to a signal that you’re cutting healthy muscle — which, if you think about it, is exactly right.

We’re at the moment when the story and the numbers start diverging in public. And when that happens, the next hard data point rules. That data point has a date: Micron, June 24.

If you want more detail I’ll leave you with my Minions, and if you enjoy the newsletter, you know that commenting or sharing is what helps me most to reach more people.

THE BOMBSHELL OF THE WEEK

SpaceX buys Cursor for $60 billion to lock down the AI coding monopoly

Days after going public, SpaceX spent the equivalent of an entire nation’s GDP to buy the code editor used by the developers who would build its competition. The move is so big it’s hard to process, so let’s walk the timeline.

April 2026: SpaceX and Anysphere (Cursor’s parent) announce a conditional agreement. The structure said it all: SpaceX reserved an option to pay ~$10 billion for a partnership, or acquire the entire company for $60 billion later in the year. That their own legal team negotiated a separate regulatory termination clause indicates they were internally treating antitrust scrutiny as a real threat, not a formality.

Monday, June 16: SpaceX signs the Agreement and Plan of Merger to acquire Anysphere in an all-stock deal with an implied equity value of $60 billion. Closing expected in the third quarter of 2026, subject to regulatory approvals. The announcement lands just days after SpaceX’s historic IPO.

Why this price. Cursor is not just any editor: it’s the interface for more than a million paying developers, over 50% of the Fortune 500 as customers, and roughly $2.6 billion in annualized revenue with enterprise sales skyrocketing. Its ARR has multiplied at vertiginous speed. But what SpaceX is buying isn’t the revenue: it’s the proprietary code data those developers generate, and the distribution channel to reach them.

The real strategic logic. SpaceX absorbed xAI at the start of 2026; Grok trailed notoriously behind Anthropic and OpenAI in the enterprise software development market. The play is textbook: take the product that already owns the developer relationship and migrate it from dependence on third-party models (Claude, GPT) to Grok running on in-house infrastructure. Buy the demand to feed the supply that doesn’t yet compete.

The drama the headlines aren’t naming. On Hacker News, Cursor’s own users are up in arms over three concrete fears that no mainstream article articulates clearly: privacy of proprietary code under a company with active military contracts; the historical slowdown of the roadmap in tools absorbed by giants; and the loss of model optionality, because if SpaceX prioritizes Grok, anyone who depends on Claude or GPT inside Cursor loses the choice they take for granted today.

The signal: vertical integration of the AI stack has reached the developer layer. Whoever controls the editor controls the most valuable training data in existence — real code solving real problems — and the channel to whoever writes it. They didn’t buy a tool; they bought the moat. And the regulatory termination fee SpaceX bothered to negotiate says they themselves know someone in Washington is going to look at this under a microscope.

POWER MOVES

Brian Hall joins Mistral: the Azure guy now sells the rival that runs on Azure

Brian Hall — a veteran with stints at Microsoft, Amazon and Google — joins Mistral as CMO, as reported by GeekWire in its round of moves for the week of June 14. This is not a cosmetic hire. Hall is exactly the profile a European lab needs if it wants to stop being “the continent’s open source model” and attack the North American enterprise head-on: someone who speaks the language of Fortune 500 buyers who already have contracts with Microsoft and AWS. The irony is delicious: Mistral has a multi-year datacenter access deal with Microsoft, so the man who knows Azure from the inside now moves to selling the competitor that runs on Azure. The signal: Mistral is going after the US market for real, not as a posture.

Qualcomm courts Tenstorrent (and Jim Keller) for $8-10 billion

Reuters, sourcing The Information, placed Qualcomm in advanced negotiations to acquire Tenstorrent, the AI chip startup led by Jim Keller — the architect behind the iPhone/iPad processors and Tesla’s Autopilot hardware. The range on the table, $8 to $10 billion, represents a notable premium over Tenstorrent’s last known valuation (~$2.6 billion). Qualcomm has been trying for years to escape the smartphone monoculture; this deal would give it AI accelerators for training and inference, its own RISC-V architecture, chiplet technology and, above all, the talent and brand around the most respected processor engineer in the industry. Both companies declined to comment and, according to Reuters, the talks may not come to fruition. The signal: the AI silicon war is no longer just Nvidia versus the hyperscalers; the mobile giants are buying their way in with the checkbook.

Vlad Tenev cuts 10% the day he says “the business has never been stronger”

On June 16 Robinhood announced the layoff of some 290 employees (10% of the workforce) in an SEC filing, with charges of ~$28 million in restructuring and ~$8 million in stock-based compensation. Tenev’s memo used the phrase “Robinhood’s business has never been stronger” — against 15% revenue growth in Q1 and record trading volumes in June. What’s revealing, as I detail in REAL NUMBERS, is what the memo doesn’t say. The signal: the “great flattening” has reached a fintech in full growth, and the script for justifying it is changing in real time.

MONEY TALKS

DeepSeek closes $7.4 billion — and the Chinese state keeps the only vote

DeepSeek closed its first external funding round since its founding on June 16: ~$7.4 billion (about 51 billion yuan) at a post-money valuation above $50 billion, making it China’s most valuable AI startup. So far, the headline. The fine print is the news: founder Liang Wenfeng personally contributed ~20 billion yuan; Tencent put in ~10 billion and battery maker CATL ~5 billion, with JD.com and NetEase rounding it out. But the commercial investors received non-voting shares and a five-year lock-up, with the capital channeled through a limited partnership controlled by Liang. The only one that obtained voting equity with no lock-up was China’s state AI fund. Translation: for the first time DeepSeek accepts outside capital, and the structure hands effective control to the state over the AI that’s already gaining share in Silicon Valley.

Odyssey raises $310 million at a $1.45B valuation — and gets off Nvidia

“World models” startup Odyssey (founders Oliver Cameron and Jeff Hawke, autonomous-driving veterans) signed the biggest round of the week: a $310 million Series B at a $1.45 billion post-money valuation, led by Natural Capital with Amazon, AMD Ventures, GV (Google), EQT and IQT. The strategic detail that doesn’t make the headline: Odyssey took Nvidia money in its seed round and in this round pivoted to Amazon and AMD, committing to use AWS Trainium silicon. It’s an explicit anti-Nvidia bet reflecting the hyperscalers’ push to wean the frontier labs off Jensen Huang’s monopoly. The week, otherwise, was one of discipline: by Crunchbase’s own count, a “slow week for large deals,” with Odyssey as the only round above $200 million and the rest of the capital clustering in $100 million club deals across cybersecurity, defense and quantum.

Accenture buys industrial cyber for ~$4.18 billion

The biggest tech deal of the week wasn’t signed by an AI lab, but by a consultancy: Accenture agreed to acquire a majority stake in Dragos (valued at $3.25 billion) plus 100% of runZero and NetRise, to build an end-to-end cybersecurity platform for critical infrastructure and OT environments. Combined enterprise value: ~$4.175 billion; closing expected between August and September. That a consulting giant would pay more than $4 billion to vertically integrate industrial cybersecurity says that “critical infrastructure defense in the era of AI-driven threats” has become a board-level decision, not a one-off product purchase. And it’s a notable liquidity event for Dragos investors.

Context that frames all of the above (backdrop, not this week’s news): Anthropic and OpenAI have confidentially filed to go public at valuations approaching a trillion dollars, and SpaceX is the furthest along in its investor roadshow. What is from this week is the pattern in the deals themselves: not a funding boom, but capital concentration in a handful of verticals — AI, cybersecurity, defense, quantum — while total round volume remains well below the peak.

PRODUCT SECRETS

GLM-5.2: open weights one point behind Claude, at one sixth the cost

On June 16 China’s Z.ai (formerly Zhipu) released GLM-5.2, a 753-billion-parameter model with open weights and an unrestricted MIT license, downloadable on Hugging Face. It beats GPT-5.5 on some long-horizon coding benchmarks and lands barely one point behind Claude Opus 4.8 on FrontierSWE, with the best score among open models on the Artificial Analysis Intelligence Index. A one-million-token context window and enterprise pricing from ~$12.60 a month. The real story isn’t the model: it’s the combination of price and license. At one sixth the cost and under an MIT license, GLM-5.2 commoditizes the middle market for coding agents and squeezes Anthropic’s and OpenAI’s margin exactly where they monetize most. The counterweight: the China data-residency risk on the hosted API pushes companies to self-host the weights, not to use the API.

Anthropic’s best model is still switched off — and that’s now a product problem

Fable 5 and Mythos 5 — Anthropic’s top class: 1M context, 128K output, always-on adaptive reasoning — have been out of service all week under a US government export control directive, and as of June 18 Anthropic was still negotiating their return. For the enterprise buyer, the product lesson is brutal: a SOTA model can become unavailable overnight by political decision, not technical failure. Claude Opus 4.8 was not affected and remains operational, but the episode turns frontier model availability into a risk variable to be managed in any architecture. The signal: put this next to GLM-5.2 and you’ll understand why the self-hosted open-weights argument just won its best commercial case of 2026 — not on performance, but on continuity guarantees. (The political standoff behind the suspension, in THE DRAMA.)

The competitive subtext: constrained at the top, commoditized at the bottom

Put the two previous stories side by side. The same month the US regulator demonstrates it can switch off a Western lab’s best closed model, a Chinese lab gives away open weights one point behind. Anyone who needs availability guarantees and data control suddenly has a very serious argument for the model they can download and host themselves. The strategic irony is hard to overstate: regulatory pressure on US closed models is, indirectly, the best sales pitch for Chinese open weights.

REAL NUMBERS

185,894 tech layoffs in 2026, ~1,115 per day. Through June 19: 267 layoff events, nearly 186,000 workers, at a pace of about 1,115 per working day according to TechTimes — almost double 2025’s 564/day. 56% of the events cite AI, automation or ML as the primary cause, equivalent to ~156,000 workers across some 150 companies. The largest single event: Oracle, 30,000 positions. The implication: the companies invoking AI to cut are the same ones committing record AI capex, and they still aren’t showing the productivity return that would justify both. It’s a story-versus-reality gap that becomes board risk if Q2-Q3 margins don’t validate it.

Robinhood: what the memo doesn’t say. The 10% layoff came against 15% revenue growth in Q1 and record daily volumes in June. And Tenev’s memo deliberately omitted any mention of AI, talking about operating “lean and disciplined” and about “frontier technologies.” The implication: it’s a leading indicator that the “AI made us do it” narrative has flipped from value-accretive to value-destructive. When you lay people off at a growing business, attributing it to AI no longer signals efficiency: it signals you’re cutting healthy muscle. Watch which Q2 layoff announcements abandon the AI framing; it’s becoming a tell.

The number weighing on the whole week: OpenAI would be losing $1.22 for every dollar. This isn’t a data point dated within these seven days — it comes from analysis of the confidential filing OpenAI submitted in early June — but it’s the backdrop that gives meaning to almost everything above, so keep it handy. The paperwork points to a debut around September at a valuation of $730 to $850 billion (some analysts argue more than a trillion); revenue would have gone from ~$2 billion annualized at the end of 2023 to ~$25 billion at the start of 2026. But per the reading of the financials, the quarter showed losses of ~$1.22 for every dollar of revenue. The implication: two of the three biggest private labs, OpenAI and Anthropic, are racing to the public markets deeply in the red — and Anthropic, this very week, had its best model switched off. The loss-per-dollar ratio is the number that public-market scrutiny can turn into the grain of sand that stops the funding flywheel. (Figure not confirmed against a primary document; the rest of this section is corroborated by tier-1 sources.)

The denominator: the capex framing it all. Meta’s 2026 capex sits at $115-135 billion (~double the prior year), plus a $100 billion / 6GW deal with AMD. OpenAI’s compute networks add up to multibillion-dollar commitments with CoreWeave and Oracle, and a Stargate already planning ~7GW. The implication: record capex + simultaneous layoffs + loss-making IPOs means the whole sector is levered to AI demand materializing on schedule. Any wobble in demand hits all three at once.

THE DRAMA

Anthropic versus the White House: the standoff over the kill switch

The power conflict of the week isn’t between two companies, but between a frontier lab and its own country’s government. And it’s being negotiated in real time.

The trigger: an export control directive forced Anthropic to suspend Fable 5 and Mythos 5 for all customers — including its own employees with foreign nationality, inside or outside the US. The company chose to switch it off for everyone rather than segment by nationality. Commerce Secretary Howard Lutnick justified the action by the fear that those models could be deployed by military intelligence users in China, Russia or other “countries of concern,” following a jailbreak method that would allow the model to be used to identify software vulnerabilities.

Anthropic’s public response has been unusually frontal for a company in its pre-IPO window: it maintains that a narrow jailbreak shouldn’t justify pulling a commercial model deployed to hundreds of millions of people, that applying that standard across the industry would paralyze the deployment of any frontier model, and that it worked with the government itself to test Fable 5 before launch — with approval to deploy it. Lutnick’s directive invoked the Export Control Reform Act of 2018, the first known use of that authority over an AI model. And the industry closed ranks: more than 80 cybersecurity executives, from firms including Nvidia and Adobe, signed an open letter asking the Trump administration to lift the restrictions. On Wednesday, June 18, Anthropic’s International Managing Director, Chris Ciauri, said at a press conference in Seoul that he was “very confident that in the coming days the models will be available again,” while senior technical staff met with the Commerce Department in Washington to find a way out. The prediction markets, however, weren’t calling it that fast: they pointed more toward July for restoration.

Estimated impact and the signal: it’s the first known use of export control authority against a specific frontier model, and it sets precedent. The relationship between the big labs and the American state stopped being laissez-faire overnight. For Anthropic, the damage isn’t just lost revenue during the blackout: it’s the demonstration, on the doorstep of its IPO, that its most valuable asset can be disconnected by administrative decision. That is exactly the kind of risk an enterprise buyer — or a roadshow investor — now has to price.

THE WEEK AHEAD

Wednesday, June 24 — Micron (MU) fiscal Q3 results, after the close. The star tech print of the week. Micron is up massively on the year on HBM memory demand for AI; its results are the real-economy proxy for whether AI capex is truly converting into chip orders. A miss or soft guidance would be the first crack in the infrastructure thesis holding up the whole sector.
Thursday-Friday, June 25-26 — PCE price index (US). The Fed’s preferred inflation gauge. It sets the rate-cut narrative underpinning high-multiple AI valuations.
DOJ/FTC response to the SpaceX-Cursor deal. The regulatory termination fee SpaceX bothered to negotiate indicates its own lawyers see the antitrust risk as real. Watch for any sign of a second request.
The return (or not) of Fable 5 and Mythos 5. Anthropic said on June 18 they’d be back “in the coming days”; prediction markets leaned more toward July. If they’re restored quickly and with conditions, it sets the precedent for how a regulatory kill switch gets negotiated; if it drags, the pre-IPO opportunity cost rises by the day.
Gemini 3.5 Pro — general availability. Google promised “next month” at its May I/O; it’s still in preview as of June 21. Every day of delay is a competitive gift to Anthropic and OpenAI. (Not confirmed.)
EU AI Act — August 2, 2026 (on the horizon). Enforcement powers over general-purpose models (fines, withdrawals, information requests) activate in about six weeks. Combined with the US export control action against Anthropic, the frontier labs face regulatory exposure on two fronts just as they head to the public markets.

REFERENCES

The bombshell — SpaceX / Cursor:

Money — funding and M&A:

Product:

Power moves:

Real numbers:

Drama — Anthropic vs. the government:

The week ahead: