Product1 distinct publisher3 min readPublished
Two million Blackwell Ultra and Rubin parts, booked five months after a one million commitment. Custom silicon did not substitute for Nvidia; it added to the bill.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
Pre-booking is the mechanism worth attention. Delivery dates of 2027 and 2028 [2] mean the allocation is being decided now, two years before anything is racked, and decided in favor of the buyer with the largest existing footprint.
Nvidia is doing the same thing one layer up. It has committed $279 billion to secure supply and manufacturing capacity, against $119 billion a quarter earlier [10]: an increase of $160 billion in three months [15], and about 2.9 times the $96.2 billion of sales it reported for the quarter [16] [8]. A supplier does not prepay at that ratio because it is unsure of finding customers. It prepays because capacity is the input it cannot buy on short notice, and each dollar of it reserves production that someone downstream also wanted.
The custom-silicon reading is where operators usually go wrong. Amazon's own chip business crossed a $25 billion annualized run rate, on $225 billion of total commitments from labs including Anthropic and OpenAI [6], and AI chief Peter DeSantis has said AWS is in talks to sell Trainium to other data center operators [7]. So this is a company with a working alternative that has nonetheless taken its Nvidia commitment to at least three million GPUs [14]. The two programs are not trading against each other at the margin; both totals are rising. And if Trainium output is being marketed to outside buyers [7], it is not a reservoir of slack for anyone else either.
The package is not confined to accelerators. Nvidia says Vera CPUs go to AWS alongside the GPUs, some integrated with Rubin and some standalone, per CFO Colette Kress [11]; Amazon plans to run Nvidia's physical AI stack on its robot fleet [12]; and AWS will serve Nvidia's Nemotron models through Bedrock and SageMaker [13]. Each of those makes the eventual cost of moving the workload somewhere else higher than the chip line item suggests.
Worth being precise about what is actually established. Both supplied reports are the same TechCrunch story [18], the figures come from Nvidia's earnings call [3], and no terms were disclosed, with the "tens of billions" valuation being TechCrunch's own arithmetic from unit costs [5]. Amazon has attached no dollar figure. Nvidia's line, that demand has exceeded expectations since the earlier commitment [4], is the seller describing its own book on the day it guided to $108 billion for the third quarter [9].
One detail undercuts any sense that this is settled procurement. Rubin only began production shipments this quarter [9]. A book covering 2027 and 2028 is therefore being placed on a generation with almost no field record, which is the price of holding a position in line: you accept the generation risk to be early in the queue rather than late.
Ranked by verification strength, evidence, and original report placement.
Amazon and Nvidia announced Wednesday an expanded partnership that includes a deal to add another 2 million Nvidia GPU chips to Amazon's data centers.
The added GPUs include Nvidia Blackwell Ultra, Rubin and Rubin Ultra parts, and will go to Amazon Web Services data centers in 2027 and 2028.
Nvidia has committed $279 billion to secure supply and manufacturing capacity for current and future data-center projects, up from $119 billion the previous quarter.
The announcement was made during Nvidia's quarterly earnings call, five months after Amazon agreed to deploy more than 1 million Nvidia GPUs across AWS infrastructure starting this year.
Nvidia said in a statement that since the earlier commitment, "demand has exceeded those expectations."
Amazon said on its last earnings call that its custom chip business crossed a $25 billion annualized revenue run rate, driven by $225 billion in total commitments from AI labs including Anthropic and OpenAI.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Firm figures, one publisher, all company-sourced
Order size, part mix, delivery years, revenue, guidance and supply commitments are all specific and attributed to named executives and an earnings call, which is strong for a disclosure story. But both supplied items are the same TechCrunch report, no contract terms were released, and the headline deal value is the outlet's own inference, so nothing here is independently corroborated.
Large realized revenue, but the new order is 2027-28 paper
Adoption is genuinely measurable on the current side: $89 billion of quarterly data center revenue, a $25 billion annualized custom-silicon run rate at Amazon, and Rubin production shipments already started. The 2 million GPU expansion itself is a forward commitment for 2027 and 2028, and the Nemotron and physical-AI integrations are announced intentions without availability detail, so realized adoption trails the announcement.
Real demand, unpriced deal and 2027-28 framing overstate the present
The underlying disclosures are substantial, so this is not empty hype. Overstatement comes from three places: a 'tripled order' headline resting on volumes that arrive in 2027-2028, a tens-of-billions value that no party confirmed, and the profitable-tokens narrative offered as justification while the article itself concedes investors have yet to see compute translate into profit. The $279 billion of commitments against $96.2 billion of quarterly sales also front-runs demand that must still materialize.
Announced on the vendor's own earnings call
The expansion was disclosed during Nvidia's quarterly earnings call by the CFO, alongside guidance and a record supply commitment, and Nvidia supplied the 'demand has exceeded those expectations' line. Amazon simultaneously benefits from signalling AI capacity while marketing its own Trainium and Graviton silicon. Both parties therefore had a direct interest in the size and framing of the number, and no terms were released that would let an outsider test it.
Specific disclosures, single-publisher account
Confidence is moderate: the quantitative core is the kind of material a public company states on the record and would have to stand behind, which supports the factual claims. It is held down by the cluster resting on one publisher's report, the absence of contract terms, and the forward-dated nature of the central commitment.
build
SemiAnalysis to software teams: your token cost starts at the fab, not the price list1 distinct publisher
product
Nvidia's $96.2bn quarter is a supply story, and 2027 capacity plans should price it that way4 distinct publishers
invest
Nvidia put a number on the bubble argument: 70% growth, and it can now be scored2 distinct publishers
build
Hugging Face's $13B process puts most teams' model pipeline under a single owner2 distinct publishers
Distinct publishers with included, body-backed reporting in this cluster.
2 articles · August 26, 2026