Our call: SEC filings put the next 12–18 months of AI capex in signed contracts; whether power and delivery keep pace is still open

Today's main line

1. [Evidence update] (filed February 11, 2026; catch-up) Vertiv's data-center power and cooling orders jumped in the fourth quarter of 2025, a step change rather than a gradual acceleration

Vertiv is one of the leading suppliers of data-center power and cooling equipment: uninterruptible power supplies, power distribution and liquid cooling. Its customers include the large cloud providers. In its February filings on the fourth quarter of 2025, Vertiv reported organic orders up 252% from a year earlier and a book-to-bill ratio of about 2.9x (8-K; 10-K). A book-to-bill of 2.9x means new orders in the quarter were 2.9 times what Vertiv shipped; in a steady state the ratio sits around 1x. Backlog reached US$15.0B, up 109% in a year.

The timing is what makes this a step change. On the same basis, organic orders grew 15% and 60% in the two quarters before. The 10-K also confirms that most of the backlog is firm orders expected to ship within 12 to 18 months.

Not proven yet: the AI share of Vertiv's US$15.0B backlog isn't broken out in the filings

Verification: The 8-K and the 10-K are two statutory filings from the same period that confirm each other. The +252% includes a low base from a year earlier and the timing of large orders, so single quarters will swing. Vertiv is not a pure AI company, and its backlog includes demand outside AI. Management attributes the orders to data centers, "particularly in AI infrastructure".

Judgment update: Read together with items 2 and 3, this item forms a new call today: the debate over whether AI capital spending has peaked needs a different yardstick. Reported revenue lags, and product launches lean toward marketing. The forward-looking figures that carry legal liability sit in filings: orders, backlog, deferred revenue and long-term contract values. Power and cooling, networking and long-term compute contracts are three independent links in the supply chain, and all three jumped together in the fourth quarter of 2025 and the first quarter of 2026. On this yardstick, whether demand has peaked is no longer an open question for the next 12 to 18 months. Two questions stay open: whether power and delivery keep pace, and how concentrated the buyers are. This call sits alongside an opposing read on our watchlist, oversupply followed by a price collapse, without picking a side. The day the orders are worked through, after 2027, is the day the window for a price collapse opens.

Our read on positioning: over the next 12 to 18 months, execution bottlenecks (power, cooling, delivery) rest on firmer evidence than the direction of demand, because firm orders have locked demand in and nothing has locked execution in. For products whose launch depends on new compute, the working assumption is that delivery runs slower than the order figures suggest. What would prove this wrong (any one narrows the call): Vertiv's book-to-bill falls below 1.5x, Arista's deferred revenue falls quarter on quarter, Meta's US$15B backstop is triggered for a large amount, or any large cloud provider cuts its 2027 capital-spending guidance. Verdict date: January 2027, checked item by item against each company's fourth-quarter results. As of this issue, none of them has cut guidance. Meta instead raised its 2026 full-year range to US$125B–145B; that figure comes from media compilations and is a directional reference.

2. [Evidence update] (four filings, August 2025 to May 2026; catch-up) At the network layer, Arista's customers are paying in advance to lock in supply

The numbers first. Arista's year-on-year quarterly revenue growth bottomed at 27.5% in the second half of 2025 and rebounded to +35.1% (US$2.709B) in the first quarter of 2026 (Q1'26 8-K). Arista is the leading maker of Ethernet switches for AI data centers. Switches are bought together with GPU clusters, because the GPUs have to be networked, so Arista's revenue is a high-frequency side reading on AI capital spending.

The more independent signal is on the balance sheet. Deferred revenue, money customers have paid for goods not yet delivered, nearly tripled in 15 months, from US$1.73B to US$4.91B. Remaining performance obligations (RPO), a broader measure of contracted revenue not yet recognized, stood at US$5.3B, about 90% of it to be recognized within two years (10-Q).

Not proven yet: how much of the rebound is organic, with about US$300M coming from VeloCloud

Verification: Four 8-K and 10-Q filings, all Arista's own primary filings, confirm each other. Revenue includes about US$300M contributed since 2025 by VeloCloud, a business Arista acquired in 2025, so the reacceleration isn't all organic growth. Part of the deferred-revenue growth reflects how multiyear contracts are booked. Arista's revenue is also highly concentrated in two large cloud customers, a concrete case of the buyer-concentration risk in item 1.

Judgment update: Deferred revenue nearly tripled in 15 months and about 90% of the broader RPO is recognized within two years, so demand visibility at the network layer now reaches out about two years. For reading suppliers like this, deferred revenue and RPO are the lines to track, not quarterly revenue alone. Prepayments and RPO growth of this kind often come with longer lead times, so off-the-shelf availability of the same class of switch isn't a safe assumption for project schedules.

3. [Evidence update] (filings of November 2025 and March 2026; catch-up) In long-term compute contracts, Meta's commitment to Nebius went from US$2.9B to up to US$27B in four months

This item is best read as one order growing over time. Nebius is a listed neocloud, a cloud company that specializes in renting out GPU capacity. Meta's first order was signed in November 2025: two dedicated GPU clusters, a five-year term, about US$2.9B (6-K). Four months later, in March 2026, the relationship expanded into an agreement worth up to about US$27B in total (6-K).

The US$27B has to be split. US$12B is firm: dedicated clusters at several sites. The filing states that they are one of the first large-scale deployments of NVIDIA's next-generation Vera Rubin platform, delivered from early 2027. The other "up to US$15B" is a backstop clause. Nebius first sells the capacity to third parties, and Meta is obliged to buy whatever doesn't sell. The US$27B therefore can't be counted as firm revenue.

Not proven yet: how much of the US$15B backstop Meta pays depends on capacity Nebius can't sell

Verification: Two 6-K statutory filings, both primary documents. Before them, "Meta is buying a lot of outside compute" rested only on a reading by the industry research firm SemiAnalysis, the one source visible at the time. This item moves that reading up to statutory-filing level, at a scale nearly ten times larger.

Judgment update: The July 14 issue covered Nvidia's backstop for CoreWeave, found in an SEC filing ten months after a podcast remark that Nvidia would backstop it. What is new today is the same structure with different parties: Meta as the backstop, Nebius as the one backstopped. A guarantee to buy whatever capacity doesn't sell is becoming a recurring structure in AI compute contracts. The same structure also masks the demand signal: as long as a backstop exists, a compute provider's results can't separate real demand from the backstop taking up slack. For reading GPU-cloud filings, the share of the backstop that gets triggered and the share of unsold capacity become necessary lines, not revenue growth alone.

4. [Evidence update] (filings of February and April 2026; catch-up) The same Nebius behind the US$27B contract disclosed an internal-control red flag

When Nebius changed auditors in February, its filing disclosed that its previous auditor, Reanda, had issued an adverse opinion on Nebius's internal control over financial reporting on April 30, 2025 (6-K). The filing's words: "Reanda's report dated April 30, 2025, expressed an adverse opinion on the Company's internal control over financial reporting." Nebius's annual report at the end of April went further. It says the weaknesses continued into 2025 and those found in 2024 have not all been fixed (20-F). Nebius gives its own reason: "the material weaknesses reflected the significant growth in our business in 2025 and the substantial changes in our organization following the complex divestment transaction we completed in 2024". That transaction refers to Nebius's predecessor, Yandex N.V., being split and restructured into an independently listed neocloud. A remediation project, supported by outside advisers, is under way.

An adverse opinion on internal controls does not mean the reported figures are wrong. It means the risk that an error goes uncaught is higher than normal.

Not proven yet: whether the weaknesses get fixed; Deloitte's first internal-control opinion isn't out yet

Verification: Two independent filings, the auditor-change 6-K and the annual report, confirm each other. Both are Nebius's own legally required disclosures, not outside allegations. As of this issue, the new auditor, Deloitte, still awaits approval at Nebius's 2026 shareholder meeting.

Judgment update: The contract figures in item 3 should be read together with this one. At fast-growing GPU clouds, growth in scale is running ahead of internal-control maturity, a named governance risk in this market. The next checkpoint is Deloitte's first internal-control opinion after taking over.

5. [Evidence update] (announced December 24, 2025; checked today) Nvidia's US$20B Groq deal: the numbers hold, but legally it isn't an acquisition

Until today the deal rested only on an account given in a Stanford class by an interested party, Groq's former president, who has since joined Nvidia. Today it was checked against the announcement and press coverage. About US$20B in cash; Nvidia's largest deal ever, above Mellanox in 2019 (about US$7B); the founder and about 90% of employees moving to Nvidia; a price about 2.9 times Groq's last funding valuation of US$6.9B. All of it holds (CNBC, 2025-12-24).

The announcement carries one more layer the class account left out. Groq signed what it calls a "non-exclusive inference technology licensing agreement", and Nvidia CEO Jensen Huang said: "we are not acquiring Groq as a company". Groq has a new CEO and continues to operate independently.

Not proven yet: the 2.5x tokens-per-power claim rests only on an interested party's account

Verification: CNBC and Yahoo coverage on the announcement day and the participant's own account confirm each other, though the person giving the account is an interested party. The technical claim of 2.5 times the token output at the same power comes only from that party and hasn't been independently checked.

Judgment update: This is the second case in a row where the value of checking was finding the sentence the retelling left out; the first was the limit on the CoreWeave backstop in the July 14 issue. The structural reading: US$20B bought the technology and 90% of the people without going through a full merger process. Tracking how large companies absorb challengers means watching licensing agreements and staff-transfer announcements; searching merger databases alone misses deals like this one.

Chips & semiconductors

1. (Filed May 6, 2026; catch-up) Arm is making a complete data-center CPU itself, with Meta as launch partner. Arm's business has been licensing chip architectures for royalties, not making chips. The shareholder letter in its May filing unveils its own data-center CPU, the "AGI CPU" (Arm's official name). That puts Arm in partial competition with its own licensees, a structural turn in its business model (6-K shareholder letter). The figures come in three grades. More than US$2B is signed customer demand for fiscal 2027–28 and can be checked. US$15B is a company target. A US$100B market by 2030 is a vision and should be discounted. The same filing reports fiscal 2026 revenue of US$4.92B (+23%, a record) and data-center royalties more than doubling year on year. One observation: Meta appears for the third time in this issue, as the compute buyer in main-line item 3, as launch partner for the AGI CPU here, and as the company raising its capital-spending range in item 1. Put plainly, Meta is at once a compute buyer, a CPU launch partner and a purchase guarantor.

2. (Filed August 8, 2025; catch-up) TSMC's July 2025 revenue was NT$323.17B, up 25.8% from a year earlier. Cumulative revenue for January to July was up 37.6%. Monthly revenue filings are a high-frequency statutory disclosure particular to Taiwan-listed companies, and the most frequent primary reading anywhere in the AI supply chain (6-K). A quieter detail in the same filing: TSMC's outstanding endorsements and guarantees for its Arizona fab stand at NT$327.3B, more than 20 times the NT$14.15B for its Nanjing fab. The ratio can serve as a proxy for how capacity is being split between the US and China. Arizona is still in heavy construction, so the guarantee balance also reflects financing needs, and it shouldn't be read directly as capacity size.

Named commentary

1. Quinn Slack (co-founder and CEO of Sourcegraph, the maker of the coding-agent product Amp), on X, July 11, 2026. Slack says Amp has become about 50% cheaper on average and much better, mainly because Amp mostly replaced Anthropic's Opus with OpenAI's GPT-5.6 (@sqs, post of 2026-07-11). In plain terms, the same product swapped its underlying model and got cheaper and better at once, which shows the model is a replaceable part inside Amp. This is the CEO's own claim about Sourcegraph's product, with a marketing motive behind a price cut, so the figure reads as direction only. The claim supports a call in the July 12 issue, on NVIDIA's balance sheet as its new product: the real product is the harness, not the model. Whoever holds the harness holds the leverage to switch models.

Model watch

1. The "memory war" gets its first direct mechanism experiment this week, and the evidence tilts toward external memory. The new evidence is a controlled experiment by researcher Charles O'Neill (July 2026, arXiv 2607.11020). It is a preprint, not yet peer-reviewed, and a single controlled experiment, so the evidence should be discounted. O'Neill wrote new facts directly into a model's weights and then kept writing other content. After 20 further writes, only 1% to 46% of the original facts survived. With the same facts placed back in the prompt as context, recovery was 77% to 80%. Two more items came the same week. Deep-learning researcher François Chollet released Morpheus on July 13, a continual-learning benchmark built on a persistent simulation in which the world never resets. Startup Engram closed a US$98M seed round, and its CEO, Dan Biderman, bet the other way on Latent Space, a widely followed AI-engineering podcast: distilling knowledge into model parameters (interview, 2026-07-13). The academic precursor of that approach is Stanford's Cartridges paper of June 2025 (arXiv 2506.06266).

The lineage of this trend: the long-context race, about reading large amounts of data at once, was settled in 2024 according to an earlier call on our watchlist. The current fight is the memory war: for AI to remember across conversations, does the memory live in external notes (context engineering) or get written into the model's parameters?

Our current call, carried over from that earlier call rather than made new in this issue: unsettled, with the evidence leaning toward external memory. The weights side hasn't yet shown that facts can be written in and stay there, and that is exactly what Engram is betting will be solved. What would prove this wrong: if two or more frontier labs ship weight-level memory in production products within 12 months, this call is void and flips. Verdict date: around mid-2027. For agent products that need memory across conversations, the external-memory side currently has the firmer evidence; the parametric route, Engram's for example, is a bet on a technique not yet shown to hold.

Product moves

No product news this issue.

Macro × AI

[This week] US inflation and hiring both cooled in June, loosening the financing premise behind AI compute contracts. The data are official figures from the US Bureau of Labor Statistics, released July 14, 2026. June CPI for all items rose 3.53% from a year earlier and fell 0.35% on the month. Core CPI, excluding food and energy, rose 2.59% from a year earlier and only 0.01% on the month (CPI release). Unemployment was 4.2%, down 0.1 percentage point on the month, and nonfarm payrolls rose by 57,000 (employment report). Core inflation close to target and a gentler job market point to more room for rate cuts.

The part of the AI capital stack most sensitive to interest rates is the GPU clouds that build compute with borrowed money. The transmission point is on paper in a primary document. Nebius's US$17.4B compute contract with Microsoft states: "The obligations of the parties under the Agreement will commence upon the confirmation by Nebius to Microsoft that it has secured any additional financing required in respect of the capital expenditures under the Agreement." (6-K, 2025-09-08) Nebius has fed its contract growth with one round of equity and convertible-debt financing after another. Cheaper capital makes contracts of this sign-first, finance-later kind easier to bring into force and speeds up debt issuance. The next link is a lower chance that backstop clauses like the one in main-line item 3 get triggered.

The leading signals to watch: the pace of convertible and new-share issuance at GPU clouds; credit spreads, the extra interest a company pays over government bonds of the same term, where a wider spread means the market sees more default risk; and backstop trigger events. Both sides stay on the record. Cheap money feeds two tails at once. In the short window, demand outrunning supply forces acceleration (main-line item 1). In the long window, easy money breeds oversupply and prices collapse once the orders are worked through. The two concern different time scales and don't exclude each other; both stay on our watchlist, with no winner assumed.

From the archive

1. (Correction archive, July 4, 2026) A quotation that never existed nearly went into our evidence. On July 3, the research system copied, from a search engine's index layer, a line attributed to the International AI Safety Report. The report is chaired by Turing Award winner Yoshua Bengio and backed by many governments; the line was attributed to its February 2026 edition. The line said that the capability thresholds that trigger stronger safety protocols had been raised at least four times in 2024–2025. If true, it would have been institution-level proof that AI labs were collectively moving the goalposts. The report's website blocked fetching at the time, so under the rules the line was marked as awaiting manual checking. The next day the 6.4MB original was retrieved and searched word for word. Neither the line nor any variant of it exists in the report (report original). The search engine's AI summary layer had made it up, and it was withdrawn the same day. What the report actually records for 2025 points the other way: after their models triggered capability warnings, three frontier labs each chose to tighten safeguards as a precaution, without conclusive evidence. The lesson worth keeping: the fake quotation leaned exactly toward a judgment already on our watchlist, that competition squeezes safety, and nearly slid into the evidence unnoticed. What stopped it wasn't intuition but a fixed rule: a quotation not checked against the original can't be marked verified. Dramatic numbers in institutional reports are best treated as fake until the original PDF shows them.

Sources & accounting

The past 24 hours. Last night's 22:13 routine fetch ran on schedule, and its numbers count toward yesterday. Most of what came in this round is a one-time backfill of history across nine channels, not new events from the past 24 hours: 2,421 academic papers, 2,241 company and personal blog posts, 974 industry newsletters, 924 company filings from EDGAR, the SEC's filing system, 620 X posts, 533 Stratechery posts, 288 podcast transcripts, 134 supply-chain intelligence items and 120 papers read closely, covering July 1 to 14, 2026. Coverage: the main line comes from the 33 filings read in full in yesterday's targeted EDGAR pass. Another 80 filings from the same pass weren't finished and are being reread, so the main line reflects signals within those 33 filings, not all 900-plus. No new X posts were fetched this round; Named commentary uses material collected the day before. This morning's 08:31 routine fetch had started and isn't counted in this issue.

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 19 outside sources in the body.

— SecondSource · generated by our research system · 19 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.