GPU renter CoreWeave's SEC filing confirms a podcast's claim that Nvidia buys CoreWeave's unsold capacity through 2032; the filing adds exit clauses the podcast never mentioned

Today's main line

1. [This week] Apple sues OpenAI for trade-secret theft, and the fight is over hardware

On July 10, Apple filed suit in the US District Court for the Northern District of California, accusing OpenAI of stealing its hardware secrets (CNBC, 2026-07-10). The complaint names two specific acts. Tang Tan, OpenAI's head of hardware and a former Apple vice president, is accused of asking job candidates still working at Apple to bring actual parts to their interviews. Chang Liu, an engineer who spent eight years at Apple, is accused of keeping his company laptop after leaving and using it to download confidential files.

The two companies announced a partnership only in 2024 to put ChatGPT into the iPhone. OpenAI then moved into hardware with its US$6.4B acquisition of io Products, Jony Ive's device startup. Over the same period, Apple engineers kept leaving for OpenAI, and the relationship cooled.

Not proven yet: every allegation is Apple's own; OpenAI hasn't responded and no court has ruled

Verification: The filing itself was reported the same day, July 10, by CNBC, Bloomberg, the Washington Post, Fortune and others, so the event is cross-checked. The allegations are entirely Apple's claims. As of this issue (July 14), OpenAI had not filed a response, no court had ruled, and OpenAI's hardware launch schedule was reported as unchanged.

Judgment update: This is the first major trade-secret case between an AI giant and an incumbent platform company, and the first time competition by hiring away staff carries a legal cost. The case connects two lines on our watchlist: OpenAI's bet on its own devices, and Apple's shift from defending talent with retention bonuses to going on the attack in court.

2. [This week] One reading says GPU demand hasn't softened, and one signal needs a discount

The one-year contract price for the H100, Nvidia's previous-generation flagship, has rebounded from a low of about US$1.70 an hour in October 2025 to about US$2.35, a rise of about 38%. That is one of two compute-market data points in the July 12 edition of Exponential View, the tech-analysis newsletter written by Azeem Azhar (EV #592). The rental figure comes from Silicon Data, a GPU price-index firm. The other data point is from SpaceX's listing filing: its compute contract runs three years but can be terminated on 90 days' notice.

Not proven yet: both figures come through one relaying source, unchecked against the originals

Verification: Both numbers appear so far in this one source, and both are relayed. The rental price comes from a Silicon Data post. The SpaceX clause comes from a primary SEC filing, but it hasn't been checked word for word against the original. The direction is usable; the figures are reference only.

Judgment update: The direction is that GPU demand hasn't softened. The 90-day exit, though, lowers the quality of the long-contract signal itself: a commitment that runs three years on paper may bind the buyer for only 90 days. On this one GPU signal, the judgment stops here. Demand isn't soft, and backlog built from long contracts deserves less weight. The full implication needs item 3's masking mechanism read alongside it.

3. [Evidence update] The backstop has three limits that no retelling of it ever mentioned

The July 12 issue, "NVIDIA's new product isn't a chip, it's its own balance sheet", reported that Nvidia backstops CoreWeave, a company that rents out GPU capacity. Today (July 14) that claim was checked against the statutory filing. CoreWeave's 8-K, filed with the US Securities and Exchange Commission (SEC) on September 9, 2025, states that when CoreWeave's customers don't use all of its data-center capacity, Nvidia is obliged to buy the unsold capacity, through April 2032; the order's initial value is US$6.3B (SEC 8-K original). Until now the claim rested only on an account by venture capitalist Bill Gurley, a partner at Benchmark, on the BG2 venture podcast in October 2025 (original podcast).

Not proven yet: the US$6.3B is an initial order value, not a guaranteed purchase amount

Verification: Verified: the filing text was checked word for word, and the podcast account matches it. The check also turned up three limits the account left out. The contract has exits: either side can terminate after a default with 30 days' notice, or after bankruptcy proceedings not dismissed within 90 days. The US$6.3B is the order's initial value, not a guaranteed purchase amount. Nvidia's obligation depends on CoreWeave meeting service-delivery conditions.

Judgment update: The call that a pure GPU cloud's financial results can't separate real demand from Nvidia taking up slack has firmed up. Its basis has moved from a podcast account to a statutory filing, and the three limits don't change the conclusion: as long as the backstop obligation exists, the results can't tell the two apart. Read with item 2, a masking mechanism doesn't mean demand has already weakened. It means that when weakness does come, the market will see it later. Three variables are worth watching: whether Nvidia discloses a related contingent liability (which would mean the backstop has been triggered), renewals before 2032, and whether 90-day exit clauses spread to other long-term compute contracts.

4. [Trend watch] (catch-up from May and June) Anthropic's reported profit, its listing filing and a skeptic's reading line up on one timeline

The Wall Street Journal reported in late May that Anthropic, the AI lab that makes the Claude models, would turn profitable this quarter (the second quarter of 2026), two years ahead of plan, with expected quarterly revenue of US$10.9B and operating profit of US$559M (as relayed by Exponential View, 2026-05-23). That is an operating margin of about 5% (our calculation: US$559M ÷ US$10.9B), a thin turn to profit rather than a money machine.

Next, on June 1, Anthropic reportedly filed confidentially to go public. On June 6, Gary Marcus, a cognitive scientist, NYU professor emeritus and longtime AI skeptic, read the headlines about Anthropic calling for an AI pause the other way round. In Marcus's reading, Anthropic will keep pushing ahead and is only floating an option it doesn't intend to take. Marcus called it "an incredible, cost-free piece of rhetoric — perfectly timed for the IPO" (Marcus, 2026-06-06).

Not proven yet: the profit figures trace to the WSJ alone, and actual Q2 results aren't disclosed

Verification: The profit figures trace to one outlet, the WSJ; Exponential View is relaying. The second quarter ended June 30, and the actual figures await formal disclosure. Marcus's post has been checked and says what is quoted here, but it is a named opinion, and whether his reading holds is a separate question.

Judgment update: The main arena of competition among frontier labs is shifting from model capability to the capital markets. The reason is how three events are spaced: the profit report on May 23, the listing filing on June 1, and the lawsuit on July 10 (item 1), all surfacing within six weeks.

5. [Trend watch] (catch-up from June) The skeptics offer three checkable anchors, and they collide with hard numbers on spending

Starting June 7, Marcus posted a run of pieces collecting evidence that AI output is soaring while its value lags (Marcus, 2026-06-07). Nominal output is surging across fields: apps, books, music, papers, web pages. Yet actual book sales dipped slightly over the same period (citing Publishers Weekly), and the return on corporate AI investment is low (citing studies from MIT, McKinsey and others).

A group of mathematicians published the Leiden Declaration, an open letter that the New York Times reported on June 2. It warns that current automated techniques "can produce plausible but unreliable (or even incorrect) arguments which are difficult to distinguish from correct mathematical proofs", and that this is straining journal review.

Analyst Gerben Wierda, whose background hasn't been established in this issue's sources, estimates that Anthropic and OpenAI, which sell AI services that write code autonomously, "may actually pay $1000 for every $100 you pay them" (as quoted by Marcus).

Not proven yet: Wierda's US$1,000-per-US$100 figure is a modeled estimate, not a company disclosure

Verification: The three anchors vary in reliability. The Leiden Declaration is real: the open letter can be linked directly and had mainstream coverage, but it was cited here through Marcus, and the full declaration hasn't been reviewed. Wierda's number is a modeled estimate that carries "may" in its own wording; no company disclosed it. All three come through Marcus alone, so they can't corroborate one another.

Judgment update: The opposing evidence is on the spending side. Epoch AI, an independent research group, measures AI infrastructure investment (data-center construction, compute hardware and networking equipment) at about 0.8% of US GDP in the first quarter of 2026 (Epoch, 2026-06-12). That is one measurer's definition. Real money is going in, and the value coming out is hard to find. Both sets of evidence stay on the record, and this issue doesn't pick a side. Two signals would settle it: whether Epoch's next quarterly reading of AI infrastructure investment as a share of GDP keeps rising, and whether corporate AI-return studies of the MIT or McKinsey kind start to recover. Marcus himself grants that coding may be the one exception, and coding is exactly the arena burning the most cash (see Named commentary).

6. [Evidence update] (June event, filed this week) The export-control episode gains its missing link: Amazon raised the alarm

The July 9 issue (in its corrected second edition) covered the export controls on Anthropic's models. On June 12 the US government, citing a jailbreak technique that had been found, required that Fable 5 and Mythos 5 stop serving foreign nationals; the restriction was lifted on June 30. This week the episode was compiled into a full record, adding a link that had been missing. According to two independent sources, the party that found the jailbreak and reported it to the Commerce Department was Amazon, Anthropic's largest investor and a major distributor of its models, with US$8B invested and another US$5B committed. The first source is the July 3 episode of All-In, a podcast hosted by Silicon Valley investors (All-In, 2026-07-03); one host, David Sacks, is the former White House AI policy lead. The second is Exponential View #578, relaying the WSJ and The Information.

Not proven yet: both sources for Amazon's role are secondhand, and Sacks's inside details are his alone

Verification: Two mutually independent secondhand sources agree that Amazon raised the alarm, which gives the claim moderate strength. Sacks has a known adversarial stance toward Anthropic, and the inside details he reconstructs come from him alone, so taken on their own they deserve more caution. The restriction has been lifted (as of July 14).

Judgment update: The call that a government doesn't need to detect AI risk itself, because a commercial partner's report can set off controls, now has a concrete case. Governance leverage is spread through supply chains and capital ties: a company's largest investor can also be the trigger for its regulation. For AI products built on the models of a single cloud partner that is also the model maker's backer, that partner can set off a cut in access on its own, so backup from other models becomes a real risk item.

Named commentary

Wix CEO Avishai Abrahami (July 13, 2026, on 20VC, the venture interview podcast hosted by Harry Stebbings): by Abrahami's math, the market values Wix's core business at close to nothing. Wix is a listed website-building software company. Last year Wix paid US$80M for Base44, then a one-person startup working on vibe-coding, where users describe what they want in plain language and AI generates the whole application. In the interview Abrahami reported (20VC, 2026-07-13) that Base44 now runs at more than US$150M in annualized revenue with about 400 staff, and that Wix puts half its free cash flow into Base44 each year, more than US$200M of about US$400M.

Abrahami then laid out a harsh sum. Wix's market value is about US$2.8B against annual revenue of about US$2.1B, so the market cap barely exceeds revenue; that is the pessimistic price the market puts on old subscription software. At vibe-coding peers' valuation multiples, Abrahami says, Base44 alone should be worth about US$8B, which would mean the market values Wix's core business below zero. Abrahami hasn't named the peers or the multiples, so the comparison can't be checked, and it is arithmetic that favors his own company. He also concedes that Base44's gross margin and customer retention are clearly worse than the core business's, because AI generation and hosting cost a lot and half-finished projects are easily abandoned. Abrahami expects improvement within two years; that is a forecast, not a fact.

Set against a judgment on our watchlist: coding is currently the one pocket of value that both bulls and bears accept, but its unit economics remain unproven. Wierda in main-line item 5 estimates heavy losses from the provider's cost side; Abrahami, from the acquired company's books, concedes weak margins that he says are improving. The direction agrees and the degree conflicts. All of Abrahami's figures are self-reported and unaudited.

Model watch

Coverage note: the academic scan of the past 24 hours brought in only one new paper (the first item below); another 182 papers were held over to the next run after a software error. "No other major new paper this week" holds only within that scan.

1. [Found this week] The first independent safety evaluation of Kimi K2.5 finds matching capability and fewer refusals. The first unofficial, third-party safety evaluation has appeared for Kimi K2.5, the open model from Chinese AI company Moonshot (arXiv 2604.03121, an April 2026 number, found in the July 13 scan). The abstract-level conclusion cuts both ways. On dangerous uses, meaning chemical, biological and nuclear-weapons knowledge, K2.5's capability matches GPT 5.2 and Claude Opus 4.5, but K2.5 refuses such requests noticeably less often. K2.5 has basic cyberattack capability but falls short of top-tier autonomous attack. Moonshot released K2.5 with no safety evaluation of its own, so this is the first. This is the only source so far: one team's own results, only the abstract readable, wholly qualitative with no numbers. The author list includes researchers connected to the system that produces this newsletter; the conflict is stated under the usual rules and loosens nothing. If the finding holds up, the implication is that as capabilities converge, risk and competition concentrate in the guardrail layer. Closed-model vendors that sell safety alignment, training a model to respect usage limits and refuse dangerous tasks, as an enterprise differentiator would see that edge diluted, and companies deploying open models would have to budget for adding their own guardrails. Main-line item 6 shows the other side of the same layer: governments also intervene directly in guardrails.

2. [Evidence update] Two productivity numbers from Ethan Mollick, quoted for three years, are now traced to the original studies. Ethan Mollick, a Wharton professor, wrote a widely read May 2023 essay arguing that AI is not good software but pretty good people (original). Both of its much-repeated numbers were traced to their sources this week. The writing-productivity gain of more than 30% comes from a randomized controlled trial published in Science: 453 participants, with time on a writing task cut by about 40% (Noy & Zhang, 2023-07). Coding more than 50% faster comes from a GitHub Copilot experiment: 95 developers, measured 55.8% faster (Peng et al., 2023-02). Verified, and Mollick's retelling is conservative: he says 30% where the trial measured 40%. Two caveats stand. The Copilot study was run by the vendor itself (GitHub and Microsoft), and its confidence interval is very wide, 21% to 89%. Both numbers measure speed on assigned experimental tasks, not productivity across a whole job. With small samples and narrow tasks, neither number supports an overall productivity gain of 30% or 50% in an ROI case.

3. [Trend watch] The NBER sets the first falsifiable scoreboard for the AI bet: 2.7x. A working paper from the US National Bureau of Economic Research (NBER), not peer-reviewed (w35290), works backward from investment data. At current capital-spending levels (US$380B by the five largest US tech companies in 2025, with 2026 forecast to nearly double), the AI sector's productivity would need to rise about 2.7x cumulatively by 2028 for the investment to pay off. If that doesn't happen, the paper says, this infrastructure wave is the largest capital misallocation in history. The scope is easy to get wrong: the 2.7x refers to the broad AI sector, not the whole economy. Co-author Jessica Wachter, a Wharton professor, clarified exactly that point to Marcus in June. The working paper's exact posting date isn't in this issue's sources. This is a model derivation, not a measurement; change the assumptions and the threshold moves. Its value is that "is AI a bubble" now has, for the first time, a threshold that can be checked.

Chips & semiconductors

1. [This week] Reading GPU rental prices starts with telling three measures apart. The new number this week, from the same EV #592: the H100's spot price, the pay-as-you-go rate for immediate rental, rose only about 10% over the same period, far below the +38% of the one-year contract index in main-line item 2. What follows is a framework for reading the two together, not a new event this week. One chip in two markets shows gains almost four times apart, because the contract price reflects certainty over the coming year while the spot price reflects how much capacity sits idle today. The third measure is generation. The H100 is the previous flagship, and its price is held up by inference demand, the everyday compute spent on models answering requests. The new flagship sits on a different supply-and-demand curve, so the two can't be read as one. A headline that GPU rents rose or fell X% says little until it names which measure it uses.

Product moves

1. AWS Bedrock. Amazon's multi-model cloud platform lets companies reach several AI model makers, including Anthropic and Meta, through one interface, then connect their own data, add safety guardrails and assemble automated workflows without building the underlying layer (product page). Why it is worth watching: Bedrock is the core of AWS's strategy of selling the platform rather than betting on a single model. As models look more like commodities, the platform layer is what holds customers and collects the money. The threatened party is the model maker that wants to tie up enterprise customers directly; the more commoditized models get, the more a neutral platform like Bedrock can cut those model makers out. Who uses it: Anthropic's models reach AWS's enterprise customers through Bedrock.

From the archive

1. The claim that ChatGPT is replacing Google search got its first behavioral data: card receipts, not surveys. Coatue, a large tech investment fund, presented a measurement at its annual conference in June 2025 (made public via the BG2 podcast, 2025-06-20). Coatue processes about 100 million credit-card receipts a day and linked the point at which a person starts paying for ChatGPT to that person's Google page views. Before subscribing, a user's Google page views grew about 4% a year; after subscribing, they fell 8% a year. The shift played out over nearly two years, gradually rather than as a cliff.

Verification: This is one firm's proprietary measurement whose method outsiders can't recheck, and it is the only source seen so far. Coatue is also a major CoreWeave shareholder and has a position when it talks about AI. The value of the measurement is that it was the first to measure search substitution through consumer behavior rather than surveys: the direction is credible, and the numbers are best read conservatively. Set against today: a year later, the contest between Google's AI ad monetization and search volume is still undecided, and this curve is among the earliest behavioral evidence in that debate.

Sources & accounting

The past 24 hours. The routine fetch starting at 21:30 last night ran on schedule and brought in 13 new pieces: 7 industry newsletters, 4 podcast transcripts, 1 Stratechery post (a paid analysis newsletter) and 1 academic paper; 6 were read in full that night. On X, this round pulled 50 posts from @paulg (Paul Graham), and all 39 original posts went into analysis. Two data-processing steps failed with the same kind of software error and were fixed at 4 a.m.; the 182 academic papers affected move to the next run, and the Model watch coverage note reflects that. Separately, an overnight run cleared the June material that had piled up in one go (29 facts and 20 sources, first published May to July 2026). That is a one-time catch-up, not counted in the 24-hour figures, and each item cited from it carries its original date. For scale: since the start of the month the system has taken in 7,399 pieces in total (2,113 academic papers, 2,083 blog posts, 1,049 newsletters, 896 company filings, 1,258 other).

The sources we track. The judgments in this newsletter rest on 529 named voices on SecondSource's long-term watchlist: 305 people on X (Elon Musk, Sam Altman, Andrej Karpathy, Greg Brockman, Demis Hassabis, Jeff Dean and others), 46 newsletters (Dylan Patel, Ben Thompson, Nathan Lambert, Ethan Mollick, Andrew Ng and others), 90 podcast voices (Satya Nadella, Dario Amodei, Jensen Huang and others), 48 paper authors (Noam Shazeer, Percy Liang, Tri Dao and others), 48 personal blogs (Simon Willison, Chris Olah, Jason Wei and others), 26 companies' earnings reports and calls (Jensen Huang, Lisa Su, C.C. Wei and others) and 70 institutional blogs (NVIDIA, Google Research, SemiAnalysis, Hugging Face and others). This issue uses 20 outside sources in the body.

— SecondSource · generated by our research system · 20 sources · Got a view? Reply and tell us

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.