On August 26, Nvidia reported second-quarter revenue of $96.2 billion, up 106% from a year ago, with Data Center alone at $89.0 billion — up 117%. Operating income was $63.7 billion on a 75.0% gross margin. The company returned roughly $26.0 billion to shareholders during the quarter and guided the next one to $108.0 billion, explicitly assuming no data-center compute revenue from China at all.
Two days later, a16z announced a $1.1 billion fund to “accelerate the physical buildout of AI.” The honest comparison between those two numbers is annual rather than headline: venture put about $4.5 billion into data centers last year — 78% of all built-environment venture — against roughly $443 billion of hyperscaler capital spending, with 2026 estimates for the five largest running between $600 billion and $800 billion depending on whose forecast you take. Venture supplied close to one percent. The same pattern repeats wherever you look this week: OpenAI’s new $400 million fund sold its entire offering to a single investor — itself; Vanguard paid $4.6 billion in cash for Altruist; Stability AI’s Series B came from Electronic Arts, Sony, Universal, Warner and AMD Ventures rather than from funds; and lenders marked down $147 billion of application-software loans while leaving infrastructure credit largely alone.
This is not a story about venture being in trouble. It is a story about scale. Once a buildout consumes more capital in a quarter than the venture industry commits to it in a year, the people who used to set terms begin taking them instead. The useful question for a founder is no longer which fund leads your round. It is whose balance sheet your cost base actually sits on.
Nvidia reported fiscal second-quarter results on August 26: revenue of $96.221 billion, up 18% sequentially and 106% year over year, of which Data Center was $89.0 billion — up 117%. GAAP gross margin was 75.0%; operating income $63.7 billion; net income $59.7 billion. The company returned about $26.0 billion to shareholders in the quarter and still holds roughly $99.0 billion of unused buyback authorization. Guidance for the third quarter is $108.0 billion, plus or minus 2%, and Nvidia states plainly that it assumes no Data Center compute revenue from China in that number. Jensen Huang’s framing was blunt: “AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue.”
The figure that matters for this brief is not the growth rate. It is the denominator. One vendor booked $89 billion of buildout spend in ninety days, while venture's entire dedicated commitment to data centers last year ran near $4.5 billion — roughly one percent of what the five largest hyperscalers spent. Venture is not absent from the AI buildout. It is simply not the marginal dollar in it, and the capital that is has different instruments, longer holding periods and a different tolerance for loss.
a16z announced a $1.1 billion “Machine Age” fund on August 28 to “open the throttle and accelerate the physical buildout of AI” — chips, memory, interconnects, data centers, robots, cooling and the electrical and real-estate work that supports them. It was widely read as the firm bucking its own software thesis. The record does not support that read. In January 2026 a16z closed $15 billion across strategies including $1.176 billion for American Dynamism and $1.7 billion for Infrastructure — both larger than this week’s fund, and both raised seven months earlier. American Dynamism itself dates to 2023.
A Form D for OpenAI Startup Fund II, L.P. surfaced this week showing a $400 million offering sold in full to one investor — OpenAI itself, financed from its own balance sheet, with a first sale date of August 11. The contrast with the original fund is the point: OpenAI’s 2021 Startup Fund raised $175 million from external backers including Microsoft. The new vehicle is more than twice the size and has no limited partners to answer to.
Leveraged-loan investors are marking application-software credit down on AI-disruption risk while leaving infrastructure alone. Application software carries $147 billion in leveraged-loan outstandings and roughly 75% of all software loans in the Morningstar LSTA US Leveraged Loan Index; its average bid had fallen to 87.75 as of July 28, down 9% year to date. Infrastructure-and-data and cybersecurity borrowers were down only 6%, at 91.25. The discount between them has widened from about one point in January to three or four points. With $32 billion maturing through 2028, refinancing remains open — but at wider spreads and steeper discounts.
Nvidia has reportedly agreed to acquire Hugging Face, the most widely used platform for sharing open AI models. The Information put the figure at $12.9 billion; other outlets rounded it to roughly $13 billion. Treat the deal as unsettled: Bloomberg described the two sides as in talks, TechCrunch as closing in, and as of this week neither company has confirmed it and no signed agreement has been reported. If it completes, the vendor that sells the compute would also own the distribution point for the models that run on it.
Vanguard and Altruist announced a definitive agreement on August 26. Vanguard’s own release describes the transaction without a headline figure; Axios reports $4.6 billion in cash. Altruist — an AI-forward custody and wealth-technology platform for independent advisors — is expected to operate as a standalone business, keeping its leadership, brand and operating model. The deal advances chief executive Salim Ramji’s effort to diversify Vanguard beyond low-fee index funds, and it closes later in 2026 subject to regulatory approval.
Instinct, a personal AI assistant reachable by phone or text, raised a $250 million Series B at a $2.5 billion valuation, co-led by Index Ventures and Benchmark, bringing cumulative funding to $350 million. Founder Noah Shinn is 23 and started the company in 2025 after leaving Sierra. The valuation has climbed from about $50 million in a matter of months while the product remains in private beta, and TechCrunch has flagged that its terms grant broad rights to retain user data for model training.
Global seed-through-growth funding to space and satellite companies has reached $20.3 billion so far in 2026, already the highest annual total on record with four months still to run, per Crunchbase. US startups took roughly $12.7 billion — more than 60% of the global total — with China at just over 20% and Europe near 10%. The category now includes orbital data centers, which is the same buildout thesis relocated above the atmosphere.
Keenable launched on August 25 with $26 million, backed by Accel, to build dedicated web search indexing for AI agents rather than human browsers. The premise is that an agent querying the web has different needs from a person — structure and retrievability over ranking and presentation — and that the existing index is the wrong shape for it. It is the kind of position venture still funds well: small, early, and upstream of a workflow that does not exist yet at scale.
Stability AI set a $76 million Series B in an August 25 funding notice, naming Electronic Arts, Sony Music Group, Universal Music Group, Warner Music Group, AMD Ventures and Pacific Alliance Ventures among the investors. EA, Universal and Warner already held strategic relationships with the company. Total financing under chief executive Prem Akkaraju now stands at $232 million across two equity rounds and convertible notes. The structure addresses the category’s hardest problem, which has never been model quality so much as the right to train on and distribute the output. Note who wrote the cheques: five corporate balance sheets and a chipmaker’s venture arm, not a syndicate of funds. That is this week’s pattern in miniature.
Deep Cogito, a San Francisco post-training research lab working on reinforcement learning and self-improvement, raised a $43 million Series A on August 27 led by TQ Ventures, with Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons and Zscaler participating. Total funding is now above $56 million. The company says the money goes to expanding its research and engineering team and to scaling the infrastructure required to train frontier models. That second line is the tell: at this size, a venture round in model research is substantially a compute budget.
Ventures Platform closed VP Pan-African Fund II at $84 million on August 27, above its original $75 million target, backing seed-stage founders across Africa. Against this week’s other numbers the figure looks small, which is exactly why it belongs here: in a market without hyperscaler capex or sovereign compute vehicles, an $84 million fund is still the marginal dollar. Venture has not stopped setting prices everywhere — only where the buildout has outgrown it.
Clearlake Capital and Google Cloud formed a strategic partnership to deliver full-stack enterprise AI across Clearlake’s portfolio companies. The mechanism matters more than the announcement: a sponsor can standardize an AI stack across dozens of holdings in one decision, which is a distribution channel no individual vendor sale can match — and, for founders selling into those companies, a procurement decision made several levels above the buyer they have been courting.
The obvious read of a16z’s $1.1 billion hardware fund is that venture is rotating into the AI buildout. The firm’s own January filings say otherwise: $15 billion closed, including $1.176 billion for American Dynamism and $1.7 billion for Infrastructure. This week’s fund is a16z’s third hardware vehicle, and its smallest.
The comparison worth making is annual and like-for-like. Venture put about $4.5 billion into data centers last year — 78% of all built-environment venture — against roughly $443 billion of capital spending by the five largest hyperscalers. Venture supplied close to one percent of the buildout it is described as betting on. Last week’s issue argued that value accrues to whoever counts the usage; this week Nvidia reported what the counted thing costs to build, and that bill is paid by balance sheets, strategic investors and credit. If you are choosing an infrastructure provider this quarter, that is a procurement question, not a market observation.
The clearest signal is not in equity at all. Lenders spent this year separating the software AI threatens from the infrastructure AI needs: application software, $147 billion of leveraged loans, now trades three to four points below infrastructure credit — against about one point in January. That spread does not stay in a spreadsheet. It arrives at sponsor-owned software companies as tighter budgets and smaller teams, which is what a repricing looks like from inside.
Here is the wager, and it is free to check. Watch the next three financings above $500 million by AI-infrastructure companies — neoclouds, data-center operators, chip startups — between now and year-end. If two or more are straight venture equity rounds rather than debt facilities or strategic investments, then equity is still competitive at this layer and I have overstated the rotation. Nvidia’s 10-Q, where the partner lease guarantees sit, is the second place to look.
The larger opportunity is not to compete for the buildout. It is to build the things the buildout makes cheap. Inference that cost a fortune two years ago is a line item now, and that is a gift to companies with a specific job to do rather than a model to train. Keenable’s $26 million to index the web for agents is worth more to its category than a billion is to the buildout, and where Ventures Platform just closed $84 million, venture is still the marginal dollar. Capital has moved beyond you at the infrastructure layer; judgment has not. Belief becomes capital — but only where belief is still what is scarce.