AI Industry Daily Briefing — September 3, 2026
Google gates a cyber-capable model behind an application form, the third lab in three days; Broadcom names its customers' gigawatts and guides AI revenue to $230bn; the US writes a G20 AI consensus designed not to bind.
The Executive Read
Two things were rationed on Wednesday, and neither was rationed by a government. Google released Gemini 3.8 Flash and, alongside it, a security-specialised variant available only to organisations it vets through a new program called Fairwind — making Google the third frontier lab in three days to put its most dangerous capability behind an application form rather than a price list. Hours later, Broadcom reported $16.7 billion of AI semiconductor revenue in a single quarter, said it has already secured the supply to roughly double that business twice more by fiscal 2028, and then named on an open call how many gigawatts of custom silicon Anthropic, OpenAI and Meta each intend to stand up and when. Access to the most capable security models is now allocated by three private admission committees applying three different and largely unpublished criteria. Access to compute is allocated by whoever controls wafers, memory and — in Broadcom CEO Hock Tan’s own ordering of his constraints — “land, power, and shell.” Meanwhile in Chapel Hill, North Carolina, the United States wrote and the G20 unanimously endorsed a set of AI principles that are explicitly non-binding and explicitly opposed to creating new regulatory bodies. The rules that bind are being written in Sacramento and Brussels. The rules written multilaterally this week were designed not to.
Elsewhere: Google signed the largest geothermal power agreement on record for a data center it has not committed to building. Commerce Secretary Howard Lutnick said chip tariffs are coming and that finished data-center servers may be in scope. AI security startups took more than $600 million in two days. And the $14 billion Nvidia–Hugging Face acquisition Bloomberg said could close “this week” still has no announcement from either company.
Top AI Headlines
Google gates a cyber-capable model — the third lab in three days to do it
What happened. On September 2 Google released Gemini 3.8 Flash and, alongside it, Gemini 3.8 Flash Cyber, a security-specialised variant available only through a new vetted-access program called Fairwind. The general model shipped broadly — the Gemini app, AI Mode in Search, Google Sheets, Gemini Enterprise, and for developers via AI Studio and the Gemini API. Introductory pricing is $0.75 per million input tokens and $3.75 per million output through December 31, after which the standard rate doubles to $1.50 and $7.50.
Artificial Analysis, an independent evaluator, scored the general model at 59 on its Intelligence Index at high reasoning, up from 56, and wrote that it “costs $0.58 per Intelligence Index task, making it the cheapest model at its level of intelligence.” A caveat from The Register is worth carrying: the model consumes more tokens than its predecessors, raising per-task cost by roughly 40%. Cheaper per token is not the same as cheaper per task.
The cyber variant is the more consequential half, and almost no coverage led with it. From Google DeepMind’s own model page, it “demonstrates frontier-level performance in autonomous vulnerability discovery,” works across 20 programming languages, and generates validated fixes after finding a flaw. Google’s own figures: 86.2% pass@1 on the CyberGym benchmark against 85.6% for OpenAI’s GPT-5.5-Cyber, and a 71.0% success rate at real-world vulnerability discovery. Google also says its Chrome Security team saw the model produce “2.6 times more correct patches” than larger commercial models, and that its Cloud Vulnerability Research team used it to find a critical vulnerability in under two hours against a process that “typically takes months.” Every one of those figures is Google reporting its own or a partner’s results. None is independently verified.
Access runs through Fairwind, which Google says gives “trusted partners a critical head start against AI-driven cyber threats, while making sure models are deployed responsibly.” Eligibility covers governments and national cyber authorities, critical infrastructure operators and core technology platforms, vetted for “a proven track record of ethical operations and research.” Google does not say which governments have access.
Why it matters. This is a pattern, not a coincidence. On September 1 OpenAI released its Astra model — at what it classes as a critical cyber capability threshold — to a small group of alpha testers, and Anthropic restricted Mythos 5.1 through a Cyber Verification Program open to US organisations only. Fairwind followed a day later. The pattern is also older than three days: OpenAI’s GPT-5.6-Cyber launched around August 10 restricted to approved users. The most capable offensive-security models in the world are no longer products you buy. They are programs you apply to.
Business implication. Security teams now queue for capability rather than procure it, which is a real problem for anyone building a vulnerability-management roadmap. The labs have created a durable enterprise tier with pricing power and no published list price. And “who is in your trusted-tester program, and on what criteria” is now a question a regulator can put to three companies at once and receive three different answers — one of which does not name the governments involved.
Sources: Google Gemini API changelog · Google DeepMind · Artificial Analysis · The Register · Thurrott
Broadcom names its customers’ gigawatts, guides AI revenue to $230 billion — and the stock falls
What happened. After Wednesday’s close Broadcom reported fiscal third-quarter revenue of $29.6 billion, up 86%, with AI semiconductor revenue of $16.7 billion, up 221% year over year and 54% sequentially. Non-GAAP diluted earnings per share were $3.32. AI was 56% of total revenue; custom accelerators, which Broadcom calls XPUs, were 73% of AI revenue.
The forward numbers came on the call rather than in the release. Broadcom raised fiscal 2026 AI revenue to roughly $58 billion, and said of the following year that it has “secured the supply to again double AI revenue to approximately $115 billion,” then expects fiscal 2028 “to again double to $230 billion.” Note the phrasing — secured supply, not projected demand.
Then it named names. Broadcom serves six XPU customers building frontier models; Tan identified four. Google: “multi tens of billions of dollars of TPUs annually over the next several years.” Anthropic: 1 gigawatt in 2026, 5 gigawatts in 2027, 10 gigawatts in 2028. OpenAI: 1.3 gigawatts in 2027 and over 5 gigawatts by 2028. Meta: 3 gigawatts across three generations of its MTIA accelerator through 2028. These are Broadcom’s characterisations of its customers’ plans. None of the four has confirmed them.
The stock fell anyway. Fourth-quarter revenue guidance of $34.8 billion — 93% growth — came in roughly $230 million below consensus while the AI component beat. Analyst Cody Acree, quoted by Yahoo Finance: “The magnitude is not quite enough from a top and bottom line standpoint on the beat and raise when you have a company that is this levered to AI.” Shares fell in extended trading, with reported declines ranging from about 3% to 6.5% depending on when the quote was taken. The September 2 regular-session close was $367.24, down 0.66%.
One disclosure change deserves its own line. Broadcom gave no AI backlog figure and no bookings figure on this call. It disclosed a $73 billion AI backlog at the end of fiscal 2025 and more than $30 billion in bookings against $10.8 billion shipped at fiscal Q2. We report the omission as a fact and draw no inference about why.
Why it matters. Nearly every AI capital-spending story is written from the buyer’s side, using hyperscaler guidance — which is a forecast. This is the seller’s side, and Broadcom says fiscal 2027 is supply already secured. It is also the first time a public company has laid out on the record, by name and in gigawatts, how much custom silicon the three most significant private AI labs intend to deploy. Note the asymmetry: Broadcom disclosed Anthropic’s multi-gigawatt commitments on an open call while Anthropic itself has published nothing about the very large compute arrangements trade outlets have attributed to it.
Business implication. Two facts sit side by side and both are true. Contracted demand is real and now disclosed. And the bar for AI-levered names is high enough that a 93% growth guide missing by two-thirds of a percent gets punished — the second consecutive night of that pattern. Informational only; we make no recommendation.
Sources: Broadcom investor relations · Investing.com transcript · Benzinga transcript · Yahoo Finance
The US wrote a G20 AI consensus, and China and Russia signed it
What happened. The G20 Innovation Ministerial concluded September 2 in Chapel Hill, North Carolina — a US-hosted meeting — with a consensus statement endorsed by all members. Three deliverables were named: the Carolina Principles for Emerging Technologies, the AI Prosperity Objectives and the AI Prosperity Compact. Six pillars run through them, including one on intellectual property policies for artificial intelligence. Commerce Secretary Howard Lutnick and OSTP Director Michael Kratsios led the US delegation; Elon Musk, Jensen Huang, Sam Altman, Tom Brown and Alex Karp took part in fireside chats.
The Carolina Principles are non-binding and, per Business Standard, call for sector-specific rule-making and for avoiding the creation of new regulatory bodies for AI. Huang’s framing at the meeting: “regulate practical and actual harm, and not regulate theoretical and hypothetical harm.”
The tension inside the consensus is the story. Business Standard reports all G20 nations backed it, China and Russia included — while China simultaneously criticised US AI firms, calling their models “distorted,” and objected to restricted access to advanced models. That last objection is a direct response to the vetted-access programs above.
Separately, and reported rather than confirmed: Axios reported on September 2 that Commerce and OSTP are “jockeying for influence” over AI policy, including over a proposed FINRA-like self-regulatory body and a replacement for Biden-era chip export restrictions. The White House denies it — “This is fake news… The President’s team is working closely together.”
Why it matters. This is the first multilateral AI document the United States has driven rather than declined. Washington skipped the November 2025 Johannesburg G20 summit and did not join its AI declaration, which framed AI as a “public good” requiring collective governance. Chapel Hill inverts that posture in the same forum, with the US holding the pen.
Business implication. The shape is loose globally, tight locally. Chapel Hill went deliberately light while Brussels runs a four-month compliance clock on ChatGPT’s designation as a very large online search engine and 26 AI bills sit on California’s governor’s desk. Non-binding principles preempt none of that. What they provide is a citable multilateral text industry can bring to every domestic fight for the next two years. The intellectual-property pillar is the detail to watch, arriving in the same fortnight Sony and Warner units sued Anthropic over training data.
Sources: The White House · Business Standard · Axios
Lutnick: chip tariffs are coming, and data center servers may be in scope
What happened. In a CNBC interview in Washington on September 2, Commerce Secretary Howard Lutnick described the coming semiconductor tariff policy in one line: “If you build here, you don’t pay, but if you don’t build here, expect to pay to enter the greatest market in the world.”
Per Korean coverage of the interview, the scope under consideration extends beyond chips to finished goods containing them — laptops, gaming consoles and data center servers. Country-specific rates and quotas are being weighed, modelled on the administration’s pharmaceutical approach. No implementation date was given and no rate has been published. Samsung Electronics and SK hynix were named as directly exposed. The same day at SEMICON Taipei, Taiwan Economy Minister Kung Ming-hsin said Taiwanese companies plan at least $20 billion in additional US investment in AI and semiconductor manufacturing; no companies were named. One outlet reported $25 billion, and we use the $20 billion figure carried by Reuters and Taiwanese press.
Why it matters. The line to pull is “data center servers.” A tariff on bare silicon lands on the chip supply chain. A tariff on assembled AI servers lands on the buildout itself — the same buildout Broadcom guided to $350 billion of shipments over two years, and the one Dell is working through a reported $95 billion server backlog to serve.
Business implication. What is described is less a trade barrier than a domestic-investment forcing function with a tariff attached. For anyone modelling AI infrastructure cost, the input is a policy with no rate, no defined scope and no date — which is itself the planning problem. Track whether “data center servers” survives into a final proclamation.
Sources: Korea Times · Korea Herald · Reuters via Yahoo Finance
Model and Product Updates
Beyond Gemini 3.8 Flash, September 2 was empty. OpenAI’s newsroom carries no item dated September 2 or 3; Anthropic’s carries none; Nvidia issued no press release either day. No open-weights model was released by any lab in the window, and no major AI product launched from Meta, Mistral, Microsoft, Amazon or Apple. A quiet day is worth stating plainly rather than padding around.
What Anthropic did on September 2 was hire. Shamez Hemani joined as a member of technical staff on compute infrastructure, after nearly two and a half years at OpenAI on the Stargate compute team and five months at Meta’s dedicated compute unit. Separately, Matt Clifford — architect of the UK’s AI Growth Zones plan, former UK AI adviser and ARIA chair — also joined. Compute procurement is being staffed as a hiring problem, not only a contracting one, at the moment Anthropic’s chip supplier is describing multi-gigawatt commitments on its behalf.
Sources: DataCenterDynamics — Hemani · DataCenterDynamics — Clifford
Regulation and Policy Watch
A correction: we reported the Pentagon was appealing the Anthropic ruling
In our September 1 edition we wrote that the Pentagon was appealing Judge Rita Lin’s ruling in Anthropic v. Department of War, sourcing it to a headline we said at the time we could not read. We have since searched for a notice of appeal or stay motion and found none. We are withdrawing that line. We are also withdrawing our description of the relief Judge Lin ordered — we said she vacated the designation and enjoined enforcement, and we have not been able to verify the remedy from the opinion or from reporting quoting it directly.
What we stand behind: in late August, Judge Lin of the Northern District of California ruled the Department of War’s designation of Anthropic as a “Supply-Chain Risk to National Security” unlawful on three grounds — unlawful retaliation under the First Amendment, deprivation of Fifth Amendment due process, and arbitrary and capricious agency action. From the opinion: “Though the Department of War is undisputedly free to select the AI vendor of its choice, the evidence demonstrates that the broad measures imposed on Anthropic were illegal and baseless.” Anthropic welcomed the ruling; the Pentagon gave TechCrunch no comment. A parallel case remains pending at the DC Circuit, argued in May, with no decision issued.
Why it matters. A federal district court has held that the government retaliated against an AI company for its safety policy — a constitutional constraint on using procurement leverage to shape model behaviour, and the pressure point in every argument about whether labs relax guardrails for defence contracts. It also leaves Anthropic as constitutional plaintiff against the federal government and copyright defendant against the major record labels in the same fortnight. How durable the constraint proves depends on an appeal that, as of this morning, has not been filed.
Sources: TechCrunch · Nextgov
California: the count is 26, and two bills would charter an audit profession
We said on September 1 that “roughly two dozen” AI bills were on Governor Newsom’s desk. The Transparency Coalition’s August 31 tally puts it at 26. As of this morning he has signed and vetoed none. The deadline is September 30.
Two deserve more attention than they are getting. AB 1405 would establish an AI Auditor Registry with state oversight. SB 813 would recognise third-party AI compliance verification organizations. Signed together they would charter a licensed AI-audit profession by statute — the registry defining who may audit, the verification framework defining what the audit is for. Who is qualified to attest that an AI system is compliant would stop being a market question and become a licensing one, in the state where most of the systems are built. Any company that will need to demonstrate AI compliance in California should be watching the registry bill specifically, because it determines who can issue the attestation.
SB 1119, the companion-chatbot bill for minors, cleared the Assembly on August 31 after passing the Senate 39–0 in May. OpenAI publicly endorsed it, arguing on its own site that protections “should apply automatically” for ages 13 to 17. No other lab has taken a public position.
One caution for readers searching this themselves: a widely surfacing article headlined “Newsom vetoes AI chatbot restrictions for kids” concerns a different, earlier bill. It is not about SB 1119.
Sources: Transparency Coalition · KQED · OpenAI
Two smaller policy notes
Release cadence versus the EU AI Act. The Next Web argued on September 2 that Gemini 3.8 Flash is Google’s third Flash model in six weeks, and that models above the AI Act’s compute threshold trigger Article 52, which requires notification to the Commission within two weeks — a window shorter than Google’s interval between releases. This is one outlet’s analysis. There is no Commission statement, no enforcement action and no Google response, and we are not reporting a violation. We carry it because the arithmetic is checkable and the question applies to every lab. [The Next Web]
Copyright: nothing moved. The Sony and Warner suit against Anthropic, Dario Amodei and Benjamin Mann was filed in late August — outlets have dated it August 28, 29 and 31, and we have not read the docket to settle it. No filing, answer or scheduling order surfaced September 2 or 3. Readers should not confuse it with the older Concord and Universal Music Publishing action, whose amended complaint containing internal discovery material was filed July 22 and is circulating again as though new. The most consequential pending decision remains Thomson Reuters v. ROSS Intelligence in the Third Circuit — argued June 11, 84 days without a ruling, and the first US appellate decision on fair use in AI training. [Axios]
Emerging Startup Radar
AI security took $130 million in disclosed rounds in a day, and an incumbent bought an agent company
What happened. Three transactions on September 2, all in the same narrow segment.
HiddenLayer raised $100 million in a Series B led by Delta-v Capital, with Ten Eleven Ventures, Morgan Stanley, M12 (Microsoft’s venture arm) and Booz Allen Ventures participating. Valuation undisclosed. It protects AI models, agents and workflows from adversarial attacks and malicious code injection. Disclosed traction: annual recurring revenue up more than tenfold over the past year to the “tens of millions,” over 90% of growth from new customers, and customers spanning financial services, the Department of Defense and the intelligence community — including one “leading frontier model provider with over 700 million weekly users.”
Lasso Security raised $30 million led by ClearSky, with Entrée Capital, iAngels, Singtel Innov8, Mindset and Swish Data participating. Sources disagree on the round’s stage label, so we apply none. It sells AI guardrails and has launched an engine that runs on CPUs rather than GPUs with sub-five-millisecond verdicts. Disclosed: tens of thousands of agents covered, billions of requests monthly, revenue up more than 500% year over year with the absolute figure withheld. Named customers include BMW, eToro, Fiverr and the US Department of Homeland Security.
Palo Alto Networks acquired Console, an eighteen-month-old company whose AI agents automate routine IT help-desk work — password resets, application access, basic troubleshooting — for integration into Palo Alto’s Cortex platform. TechCrunch reports the price at roughly $500 million in cash and stock, sourced to people familiar rather than confirmed; Palo Alto declined to comment on the figure. Console’s pre-acquisition valuation was $157 million per PitchBook. Backers included Thrive Capital, DST Global, and Palo Alto CEO Nikesh Arora personally as an angel investor.
Why it matters. Add the $50 million AIR Security raised the day before and the segment took well over $600 million in two days. The growth rates disclosed — tenfold at HiddenLayer, above 500% at Lasso — are off small bases but they corroborate each other. This is the containment layer being funded at precisely the moment the labs finished building the gate.
Business implication. Agent security is a budget line now, not a research topic, and a platform incumbent has paid a reported 3x step-up for a company that did not exist two years ago — the shape of a category being consolidated before it matures. Anyone running agents in production should expect to be sold this category within two quarters, and should ask vendors for third-party validation rather than self-reported figures.
Sources: TechCrunch — HiddenLayer · SiliconANGLE — Lasso · TechCrunch — Console
The day’s largest round disclosed no revenue
Wonderful, an Israeli-Dutch company founded in 2025, raised a $550 million Series C led by Insight Partners at a $5 billion valuation, with Salesforce investing for the first time alongside Index Ventures, IVP, Vine Ventures, 9Yards and Bessemer. That is more than double the $2 billion it carried at a $150 million Series B in March 2026. It sells a platform coordinating AI agents, workflows and applications against a company’s own data and systems, and reports operating in more than 35 markets with 650 employees. No revenue or customer count was disclosed.
Note what is absent. Several smaller rounds this week disclosed revenue or growth rates; the largest did not. Buyers evaluating agent platforms should ask for production references rather than reading valuation as validation. Salesforce joining as a first-time investor in a company adjacent to its own agent products is the detail worth watching.
Also on September 2: Aslan Protects raised $20.8 million in seed funding co-led by Khosla Ventures and XYZ Venture Capital, for a cyber-defence platform deploying what it describes as human-supervised autonomous AI operatives for national security and federal law enforcement customers; no traction disclosed. And Adobe acquired Rilo, an Indian market-intelligence startup founded in 2025 with six employees, for undisclosed terms — Adobe confirmed the deal and declined further comment, and Rilo will shut down and stop serving existing customers.
Sources: Business Wire via Yahoo Finance · TechCrunch — Wonderful · FinSMEs · TechCrunch — Rilo
AI Infrastructure and Market Signals
Google signed the largest geothermal agreement on record — for a data center it has not committed to building
What happened. On September 1 Fervo Energy and Google signed a 396-megawatt power purchase agreement from the Cape Station enhanced geothermal project in Beaver County, Utah. Enhanced geothermal drills into hot rock that lacks natural water flow and creates the flow artificially, which makes geothermal viable outside the few places with natural hot springs. The agreement carries an expansion option to roughly 996 megawatts — nearly a gigawatt — by June 2030, with power expected online in 2028. Fervo describes it as the largest geothermal offtake in the sector to date.
The conditionality must travel with the item: the agreement powers a potential Google data center, and per DataCenterDynamics the building itself remains conditional on engineering feasibility, regulatory approvals and commercial viability. The power is contracted. The data center is not. Google has backed Fervo since 2021 and signed a 115-megawatt agreement with Fervo and NV Energy in June 2024; Fervo went public in May 2026 at a $10 billion valuation.
Why it matters. Firm, always-on clean power for data centers has been framed almost entirely as a nuclear question, and there were no nuclear or small-modular-reactor agreements in this window. The answer arrived from a different direction — enhanced geothermal, at near-gigawatt scale, from a company that listed four months ago, on a tariff structure that has already survived a state regulatory process.
Business implication. Read it against Hock Tan naming “land, power, and shell” as his first constraint the following day. The buildout is power-gated, and geothermal has moved from pilot to procurement option. What it has not done is prove delivery: first power is 2028.
Sources: Fervo Energy investor relations · DataCenterDynamics
TSMC’s own equipment forecast nearly doubled in six months
At a SEMICON Taiwan panel on September 2, TSMC Senior Vice President Cliff Hou disclosed how the company’s internal equipment-purchase forecast has moved against a late-2025 baseline: 1.0x in late 2025, 1.5x by the first quarter of 2026, 1.9x by July 2026. Separately, TSMC said it will build an advanced packaging hub at Baipu Industrial Park in Kaohsiung — a testing, co-development and training facility rather than a production fab — and expects CoWoS advanced packaging capacity to grow at a compound annual rate above 80% through 2027. No investment figure was disclosed.
A forecast revision disclosed by the company doing the buying is a cleaner read on the slope of AI demand than any analyst estimate, and TSMC would know first. It corroborates Broadcom’s supply commentary from an entirely independent direction — and the choice to build a validation site rather than more wafer capacity says something about where the integration bottleneck now sits.
Sources: Focus Taiwan — SEMICON panel · Focus Taiwan — Kaohsiung
Capacity announcements, and one company leaving bitcoin for AI
Cerebras broke ground on a 165-megawatt campus in Mikkeli, Finland with partner Compute Nordic Finland — 50 megawatts under construction now, rising to 165, at €1 billion to €1.7 billion, with closed-loop cooling and waste-heat recovery to the local district. Compute Nordic CEO Pyry Virrantaus took a direct swipe at the speculative-capacity critique: “This partnership with Cerebras is not a speculative bet on future demand—it’s a contractually committed, phased build-out.”
IREN signed an unnamed “leading frontier lab” and said it will effectively decommission bitcoin mining by the end of December 2026. Contract value, capacity and timeline were all undisclosed, and we are not speculating about which lab. IREN reported $4 billion of annual recurring revenue for fiscal 2026 and $16.6 billion of remaining performance obligations — but mining is still the majority of revenue, $578.2 million against $128.8 million from AI cloud, and capacity at June 30 was 380 megawatts of mining against 40 of AI cloud.
Also on September 2: AWS and Humain committed to 50 megawatts at the Saudi AI Zone by 2028 within a $5 billion Riyadh investment — a milestone update on a May 2025 partnership, not a new agreement — and Amazon filed for a US–Japan subsea cable at 420 terabits per second, expected 2029, the first international cable to land in Washington State in nearly 30 years.
Cerebras and IREN are worth reading against each other: one is marketing contracted, phased capacity as a rebuke to speculative buildout, the other is converting a mining fleet into AI capacity on the strength of a customer it will not name. Two very different risk profiles funded in the same segment on the same day.
Sources: Cerebras · IREN · AWS and Humain · Amazon
Data-center opposition is now about non-disclosure agreements
A September 2 feature from KCUR documents the shape of local opposition across the Midwest. Nebraska Governor Jim Pillen signed an executive order in August rescinding data center tax incentives — “We cannot foot the bill for them… we also can’t afford projects that deplete our natural resources.” The St. Charles, Missouri city council unanimously passed a one-year moratorium in August on a 440-acre project later identified as Google’s; residents could not learn the company’s identity because of non-disclosure agreements. Illinois has halted its own incentives. Gallup found in March 2026 that seven in ten Americans oppose data centers in their area.
The through-line is that NDAs, not power or water, are now a stated driver of opposition: officials cannot tell residents who the applicant is before voting, and communities are voting no on the secrecy. That connects directly to Tan naming land and shell as his primary constraint the same day. A Republican governor withdrawing incentives is also a different signal from a town moratorium, and worth tracking separately.
Source: KCUR
Public Investment Watchlist
Informational only. Artificial Record does not make investment recommendations.
Wednesday closed higher on very narrow breadth. The S&P 500 closed at 7,666.60, up 0.5%; the Dow at 53,061.95, up 0.6%; the Nasdaq Composite at 26,217.83, up 0.5%; the Russell 2000 at 2,953.17, up 1.1%. The VIX fell 6.79% to 15.23.
The breadth is the number to hold. Per Motley Fool’s arithmetic, Nvidia alone contributed 0.38 percentage points of the Composite’s gain, and excluding Meta and Nvidia the Composite would have closed down. Nvidia rose 3.21% to $224.41, Meta 2.47%, Micron 2.43%, Oracle 3.13%. Dell rose 15.81% to $492.20 — a full session after reporting a $95 billion AI server backlog, meaning the market took an entire day to price it. Against that, Palantir fell 5.81% and Snowflake 4.37% in the regular session.
Three of four Wednesday-night reporters beat, raised and sold off. The exception was Snowflake: product revenue of $1.491 billion, up 37% and a third consecutive quarter of acceleration; net revenue retention of 126%; remaining performance obligations of $9.00 billion; full-year product revenue guidance raised to $6,070 million. Management attributed roughly half the revenue acceleration to AI products — a specific attribution rare enough to be worth quoting. The stock rose more than 22% after hours. HPE and NetApp both beat on both lines and raised full-year guidance, and both fell after hours. Two consecutive nights of beat-and-fall describes how high the bar sits for AI-levered names, not a verdict on demand.
Rates. The federal funds target range is 3.50%–3.75%, held in July with three dissents favouring a hike; the FOMC meets September 16. Futures were pricing roughly a 66% chance of a 25-basis-point increase as of the most recent anchored reading we could obtain, which is now several sessions old — we have not been able to reach the CME’s own tool and do not treat secondary summaries as current. The driver remains Chair Kevin Warsh’s August 28 Jackson Hole remarks: “We must be confident that underlying inflation is moving to our objective, clearly and at sufficient speed. Otherwise, we have work to do.” The ten-year Treasury sits near 4.78–4.81% depending on source and timestamp, close to its 52-week high. Brent crude was around $95 on September 3, up roughly 20% on the month.
One piece of context for the Broadcom reaction: in late August, Google committed to up to $120 billion of revenue through fiscal 2033 with Marvell, granting Marvell a $12.2 billion stock warrant. Marvell fell more than 8% on the announcement because the largest revenue contribution does not arrive until fiscal 2029. Reuters cited custom-chip competition explicitly as the pressure on Broadcom’s multiple.
Sources: Motley Fool · stockanalysis.com
Forward-Looking Angle
Earnings tonight. Zscaler reports after the close, call at 4:30pm Eastern, consensus $1.09 per share on $876.96 million. Also after the close: UiPath, Samsara, Guidewire, DocuSign, Ambarella and Asana. Given two consecutive nights of beat-and-fall, the interesting variable is not whether these companies beat but what the market does with it.
Nvidia and Hugging Face: still nothing. Bloomberg reported on September 2 that Nvidia was nearing a roughly $14 billion deal to acquire Hugging Face “this week.” As of this morning neither company has announced anything. Nvidia’s newsroom carries no release dated September 2 or 3, Hugging Face has published nothing, and repeated searches return only secondhand coverage dated August 26–27 or the September 2 Bloomberg headline. Every headline in circulation still says “nears,” “closes in on” or “report says.” We expect much of the category to write this up as done. It is not. If it signs it is the largest story of the week — the dominant AI chip vendor acquiring the dominant open-model distribution platform, at roughly three times the $4.5 billion valuation Hugging Face carried in 2023, raising immediate neutrality questions about a platform that hosts the weights of Nvidia’s customers and competitors alike.
Anthropic’s IPO filing is reported for as early as Tuesday, September 8, per IPOScoop, with a Nasdaq listing expected. Every valuation figure circulating for the company is unverified and the figures do not agree with one another. We will report the prospectus when it is public and not before.
Watchlist
Whether Google, Anthropic, OpenAI or Meta confirms, disputes or declines to comment on the gigawatt figures Broadcom attributed to them. A denial would be a significant story; silence is itself informative.
Broadcom’s “secured the supply” claim for fiscal 2027. Tan named memory sourcing and substrate availability as constraints. Memory-maker commentary either corroborates that or does not.
Who gets into Fairwind and the equivalent programs at OpenAI and Anthropic, and on what criteria. Google has not said which governments have access.
Newsom’s decisions on the 26 California AI bills, deadline September 30 — particularly AB 1405 and SB 813 together, and whether any lab other than OpenAI takes a public position on SB 1119.
Whether the semiconductor tariff proclamation lands, and whether “data center servers” survives into its final scope.
The Third Circuit in Thomson Reuters v. ROSS Intelligence — argued June 11, 84 days and counting, the first US appellate ruling on fair use in AI training.
Whether the Pentagon appeals Judge Lin’s ruling, and what relief she actually ordered, which we have not been able to verify.
IREN’s unnamed frontier lab customer, and whether a company still earning three-quarters of its revenue from bitcoin mining exits that business inside four months as stated.
The Utah data center the Fervo agreement is meant to serve. The power is contracted; the building is still conditional.