❯ Nvidia closes up 8.74%, adding $442 billion in market value in a single day

[post-earnings jump] Nvidia closed Thursday up 8.74% at $227.98, adding $442 billion in market value in one session — the second-largest single-day gain by any stock on record. The trigger was its just-reported second quarter: revenue of $96.2 billion, more than doubling year over year and beating the $92.2 billion analysts expected, with non-GAAP earnings of $2.22 a share, above the $2.06 to $2.09 range forecast.

[the guidance did it] What actually turned positioning was the next quarter. Nvidia guided to $105.84 billion to $110.16 billion in third-quarter revenue, implying 85.7% to 93.2% growth against $57.01 billion a year earlier. The worry going in had been that hyperscalers were slowing purchases and AI demand had peaked. Jensen Huang told the call that AI infrastructure buildout “is at full steam,” and addressed the “circular financing” criticism head-on.

[who has to rethink] One earnings report pushed the skeptics back down the whole AI trade. Hedge funds short AI capex have to redo their assumptions first: if hyperscaler orders hold into the first half of next year, the peak-demand premise collapses. Index funds cannot sidestep it either — Nvidia alone carries close to an 8% weight in the S&P 500, so its swings rewrite the risk exposure of passive money that never chose to buy AI at all.

▪ SIGNALNvidia’s quarterly guidance now functions as a macro data release; one company’s earnings call sets the positioning for everyone else.

❯ Anthropic weighs letting insiders sell in its IPO, with lockups longer than 180 days

[deal structure] Anthropic is working on a plan to let existing shareholders sell some stock in its IPO while considering lockups longer than the standard 180 days for at least some holders, The Information reported. The prospectus is set to go public after Labor Day, targeting a listing in late September or early October. The plans are not final.

[breaking from SpaceX] The structure would separate Anthropic from the year’s two other big tech listings: neither SpaceX nor Cerebras let existing stockholders sell in their IPOs. Opening up secondary sales adds float, and the longer lockup is the offset — early investors and employees get liquidity in the offering, while pressure on the open market gets pushed out at least another quarter. Anthropic’s raise is reported to exceed the $86 billion SpaceX brought in.

[who recalculates] Private-market buyers of Anthropic stock have to redo the math: if shares can be sold in the offering itself, the discounts priced against a closed exit go away immediately. Employee option economics shift too, with a lockup past 180 days leaving a gap between paper wealth and usable cash. The tightest spot is the underwriters’ price range, where new issuance and insider supply arrive together and thin the cushion a normal book would have. Watch how many lockup tiers the prospectus defines, and who falls into each.

▪ SIGNALOpening secondaries while extending the lockup are two ends of one decision: liquidity now, but no immediate exit.

❯ Anthropic discussed buying chip startup MatX for $7 billion; talks are dead

[in-house silicon stalls] Anthropic held talks to acquire AI chip startup MatX for roughly $7 billion, aiming to bring custom hardware design in-house, Reuters reported. The talks are no longer active and have shifted to a partnership discussion. MatX is now seeking new capital at a valuation of about $4 billion. Reuters could not determine why the deal fell away.

[what MatX is] Founded in 2022 in Fremont, California, MatX was started by Reiner Pope and Mike Gunter, both veterans of Google’s TPU team, and builds chips tuned specifically for large-model training and inference. The company closed a $500 million Series B; its 2024 Series A carried a post-money valuation of only about $300 million. The jump from $300 million to a $7 billion offer says buyer urgency around custom silicon rose faster this year than the company itself did.

[the cost structure] Designing chips is expensive and slow, and buying a team is the shortcut. With the deal dead, Anthropic’s long-run inference costs stay tied to outside suppliers — awkward for a company preparing an IPO and explaining its gross margin path to investors. MatX’s $4 billion standalone round offers the other answer: while every large buyer wants custom silicon, independent chip companies have no shortage of acquirers, only a shortage of reasons to stay independent.

▪ SIGNALA $7 billion acquisition failed while a $4 billion independent round did not — pricing power in custom silicon is moving back toward the seller.

❯ NYT: Meta internally projected spending up to $10 billion a year on Anthropic’s models

[criticizing and buying] Meta has internally projected annual spending of as much as $10 billion on Anthropic’s AI models even as Mark Zuckerberg publicly criticizes the rival, The New York Times reported. Around the same period the two are negotiating a deal running the other direction: Meta leasing compute to Anthropic, also in the range of $10 billion over two years.

[two opposing cash flows] The compute-leasing talks surfaced only in July and remain early, according to reports. One flow has Meta paying for Anthropic’s model capability; the other has Anthropic paying for Meta’s data centers. Zuckerberg frames external leasing as a hedge on self-built capacity — if internal demand does not fill the racks, rental revenue covers the gap. That is also why the public posture and the purchasing decisions can run in parallel: the criticism is about strategy, the purchase is about capacity.

[who recalculates] Enterprise buyers can read a blunt signal here: even the company most aggressive about training its own frontier models is budgeting billions for someone else’s, which makes “build fully replaces buy” a shakier economic case than it looks. Anthropic’s revenue mix gets rewritten in the process — one of its largest customers is the competitor attacking it in public.

▪ SIGNALWhen public criticism and a ten-figure purchase order coexist, believe the purchase order.

❯ DeepSeek to raise $7.4 billion at a $74 billion valuation, aimed at compute buildout

[round restarts] DeepSeek is close to raising about $7.4 billion from new and existing investors at a post-money valuation near $74 billion (roughly 500 billion yuan), with proceeds going to R&D and compute infrastructure, the Wall Street Journal reported. The round was paused last month, reopened in August, and signing is expected within the month.

[where the money goes] Part of the capital is earmarked for self-built data centers, including one large facility in Inner Mongolia expected to cost billions of dollars. In June, DeepSeek already raised more than 50 billion yuan at a post-money valuation around 450 billion yuan — two rounds essentially back to back, with the valuation up roughly 11% in a little over two months. The company is also in early preparation for a listing on Shanghai’s STAR Market, with an internal target of filing this year.

[from model shop to heavy assets] A company known for cheap training is now pushing serious money into racks, which puts a deadline on the open question of whether “cheap” was an algorithmic result or a stage that simply preceded owning capacity. The real squeeze lands on the key-account lists at Chinese cloud providers — DeepSeek is moving from renting compute to building it. Their window runs until that Inner Mongolia site comes online.

▪ SIGNALOnce the racks are self-built, DeepSeek’s cost advantage stops being an algorithm question and becomes a capex question.

❯ Filing shows Nvidia still owes $18 billion in equity investment commitments this year

[the investor role] Nvidia’s latest quarterly filing shows that as of July 26 the company had $18 billion in equity investment commitments due over the remainder of the fiscal year, with $25 billion committed across all future periods. Its holdings in private companies carried a book value of about $47.9 billion over the same period.

[how big it got] Per the quarterly filing, the comparable figure in early 2020 was under $100 million. Including public-market positions, Nvidia’s total equity holdings have reached roughly $95.6 billion, and investment gains alone contributed $7.8 billion this quarter. Counting supply commitments, cloud purchases, data centers and capex together, future obligations total about $366 billion, spread across model developers, infrastructure financiers and a bench of private companies.

[the circular-financing math] These are the numbers behind the “circular financing” criticism: Nvidia invests in customers, and the customers buy its chips. Auditors and regulators will be watching revenue recognition next. Investors long Nvidia have to accept something else — what they own is no longer just a chip company, but a nearly $100 billion portfolio bet almost entirely on AI.

▪ SIGNALA $47.9 billion private book turns Nvidia’s quarterly filing into a valuation sheet for the AI private market.

❯ Uber’s weekly AI agent requests up 9.4x while total AI spending has been flat since April

[usage up, bill flat] Uber disclosed that weekly AI agent requests have grown 9.4x since February while total AI spending has held roughly flat since April. The turning point came in the first quarter, when the company burned through its entire 2026 AI budget in one go and its chief technology officer publicly questioned whether the money was well spent.

[where the efficiency came from] On the same model, cost per 1,000 requests is down nearly 34% from its April peak and cost per session is down 52% from its June high. The CTO changed his tone in August, saying “the so-called tokenmaxxing era is coming to an end,” with the next phase about using tokens more efficiently. AI agents now account for more than 70% of code-change submissions at the company.

[how budgets get set] The curve gives enterprise technology leaders a citable precedent: a 10x jump in agent usage does not have to mean a 10x bill, with model routing, context trimming and caching sitting in between. It also puts a question mark over the revenue models of cloud and model vendors — if large customers’ usage growth gets absorbed by falling unit costs, the per-token growth curve is less steep than it appears.

▪ SIGNALUsage up 9.4x with a flat bill is the most valuable engineering result enterprise AI produced in the past year.

❯ SoftBank in talks for a majority stake in humanoid maker 1X at a $6 billion valuation

[control, not participation] SoftBank is in talks to buy a majority stake in OpenAI-backed humanoid robot maker 1X at a valuation of about $6 billion, with terms still subject to change, The Information reported. The price sits well below the $10 billion valuation 1X sought last year, when it planned a $1 billion raise and landed less than half of it.

[product and orders] 1X’s flagship NEO is built for household chores, priced at $20,000 outright or $499 a month, and the company says its first-year capacity of 10,000 units sold out in five days — though no robot has actually reached a customer yet. For SoftBank this follows a pattern: last year it agreed to buy ABB’s robotics business for $5.4 billion, and when it first engaged 1X in 2023 the pre-money valuation was just $375 million. In three years the price on SoftBank’s books rose more than tenfold while delivery stayed at the reservation stage.

[from investing to consolidating] Moving from investor to controlling owner says Masayoshi Son wants the production line itself rather than a financial return. The valuation anchor for humanoid robotics shifts down with it — $10 billion could not get funded last year, $6 billion buys the whole company this year, and the discount the private market applies to “not yet delivered” is now explicit. Meanwhile, Tesla is still working toward a consumer launch for Optimus.

▪ SIGNALThe distance between sold out and delivered is the entire difference between a $10 billion valuation and a $6 billion one.

❯ OpenAI tests a persistent Codex mode that works across sessions until put to sleep

[spotted in the code] OpenAI is testing a Persistent mode for its coding agent Codex, letting the agent keep working across sessions and create its own follow-up tasks until a user explicitly puts it to sleep. Wired found the code in the repository for Codex’s command-line version. OpenAI says it has no near-term plans to ship the feature and that it remains research-stage.

[how it differs] Existing modes stop after minutes or hours even when a task is unfinished. According to the code Wired reviewed, once Persistent mode is selected Codex will “continue working until put to sleep,” using past interactions and knowledge of the user to decide what to do next. The code appeared in the command-line repository in August and is not enabled in any shipped build. That CLI has historically been the testbed, and most features there eventually reached the desktop and web clients.

[who should plan ahead] Enterprise security teams need rules in place first: an agent that runs continuously and assigns itself work needs permission boundaries and audit logs designed for a resident service, not a one-off session. The tighter squeeze is on per-session cost models, which an agent that never clocks off would invalidate entirely.

▪ SIGNALGoing from “run a task” to “always on” changes both the billing unit and the permission model for agents.

❯ Shein prices Hong Kong IPO near $6.20, raising $1.73 billion at a $26.5 billion valuation

[priced at the midpoint] Singapore-headquartered Shein is set to price its Hong Kong IPO at HK$48.56 a share (about $6.20), near the midpoint of its marketed range, raising roughly $1.73 billion at a valuation of about $26.5 billion, Reuters reported.

[distance from the peak] That is roughly one quarter of the near-$100 billion private peak Shein carried in 2022, and well under the $66 billion valuation from its 2023 round, per earlier reports. The company had planned a London listing, and turned to Hong Kong after that effort stalled in 2024 — more than two years end to end. Shein says about 80% of proceeds will go to technology and global brand expansion, with the rest to working capital.

[who is watching the price] Chinese tech companies queued up for Hong Kong will treat this as the reference case: even priced at the midpoint and steeply discounted from the peak, a $1.73 billion book still filled. Early private shareholders meanwhile absorb a loss that is now publicly confirmed. The tighter constraint is the price range for the deals that follow, since first-day trading will decide how aggressively the next issuers dare to mark themselves.

▪ SIGNALFrom $100 billion to $26.5 billion, Shein’s pricing settles four years of private-market inflation in a single print.