Nvidia's $279bn of reserved capacity now sets the price of everyone else's compute
Forward supply and manufacturing commitments nearly tripled in six months to $279bn, which means Nvidia is buying memory and fabrication capacity years ahead of orders against its own forecast of what everyone else will spend.
The Board Room · Leadership desk
What happened
- Nvidia reported $96.2bn of revenue for the quarter, up 106% from a year earlier, with data centre revenue growing 117% to $89bn.
- Committed future supply and manufacturing capacity stood at $279bn at the end of July, against $119bn three months earlier and $95.2bn at the end of January.
- The commitments are scheduled at roughly $92bn for the remainder of the current fiscal year, $87bn for fiscal 2028 and $88bn for fiscal 2029.
Why it matters
- What it limits constraint Capacity held under a multi-year reservation cannot be bid for by a rival accelerator vendor or an in-house silicon programme at the same price, so a buyer's 2028 options narrow before that buyer starts planning.
- Who's exposed exposure The right to cancel, reschedule or adjust protects the party that wrote the order; a customer whose delivery slot moves has no equivalent clause, so schedule risk lands downstream.
- Who has to decide decision Anyone sizing compute for 2027 and 2028 is now choosing between committing early to hold a queue position they do not control and staying flexible for whatever clears after the reservations are filled.
- What it sets up precedent Once a supplier books three years of memory and fabrication capacity at this scale, pre-booking becomes the expected behaviour, and buyers who cannot write those cheques inherit the residual.
Next story · Invest · Oct 2
Nvidia's $235 billion buyback would retire about 4% of its shares at the record priceConfirmed5 publishers
A supplier that reserves memory and packaging years ahead of shipment stops being a price taker in its own supply chain and becomes the party that decides what is left for anyone else to buy. Jim Osman's account of the quarter for Forbes puts the mechanism plainly [16]: high-bandwidth memory is one of the binding constraints in AI infrastructure, and supply a competitor cannot reach is itself an advantage, because it lets Nvidia ship more systems and lowers the chance that a shortage opens a door for alternative technology [7]. Read from the buyer's side, that is a statement about what a 2028 procurement plan will be permitted to contain.
The commitment line is worth doing longhand. Going from $119bn to $279bn is $160bn of fresh obligation booked inside three months, an average above $50bn a month [13]. The three named windows account for $267bn of the total, which leaves roughly $12bn scheduled outside them [17]. Measured against the January figure of $95.2bn, the number is about 2.9 times larger in six months [14]. What cannot be computed is the per-quarter outlay, because the source does not say how many quarters remain in the fiscal year carrying the $92bn [12].
A skeptic reads the cancellation language and stops there. Osman reports the commitments cover memory and manufacturing capacity for current and future generations of Nvidia's data centre systems, that some can be cancelled, rescheduled or adjusted, and that it would be wrong to call the figure debt [6]. That is fair, and it is why the number does not belong in a leverage screen. It does less for a compute buyer, because the flexibility belongs to whoever wrote the order: a slot Nvidia can release is still a slot nobody else could bid for while it was held, and a date Nvidia can move is a date its customer cannot.
The demand evidence in the quarter supports the bet so far. Data centre revenue of $89bn is 92.5% of the $96.2bn total [11], the Vera Rubin platform is in full production with systems running at Google Cloud, Microsoft Azure, Oracle and CoreWeave, and orders are arriving from AI labs, enterprises, sovereign customers and industrial users as well as the hyperscalers [10]. Guidance of about $108bn for the current quarter, and roughly 70% growth for the next fiscal year, sits ahead of what Wall Street had modelled [2].
What changed is which risk Nvidia carries. Long lead times have always forced semiconductor firms to commit to foundries, memory and packaging before a finished product reaches a customer [15]; the scale here turns a supply problem into a forecasting problem, and Osman's point is that reserving years in advance obliges Nvidia to be accurate about eventual customer need rather than merely responsive to orders in hand [8]. Nvidia's own caution names the failure mode: customers can delay new architectures, struggle to finance infrastructure, or adopt more gradually than expected, pushing revenue out while some commitments stay in place [9]. Osman sees nothing in this quarter to suggest that is happening [4].
The $96.2bn and the $108bn guide are this quarter's facts; the $88bn pencilled into fiscal 2029 is a claim about the back half of the decade [1][2][5]. For a budget owner, the second number outlasts the first. While those reservations hold, the availability and price of accelerators are set by one company's read on demand three years out, and the one scenario that loosens that grip is the scenario where the forecast proves too high and reserved capacity comes back to the market at a discount.
What to watch
- Whether the commitment figure climbs again next quarter, and whether the fiscal 2029 tranche moves in either direction.
- Any breakdown of the $279bn into memory versus foundry and packaging, and how much carries a penalty-free cancellation right.
- Whether large compute buyers start disclosing their own multi-year supply prepayments in response.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence72
- Adoption74
- Hype gap+22
- Incentives64
- Confidence68
Perspective Coverage
4 publishers- Builder
- Builder 14%
- Operator
- Operator 21%
- Investor
- Investor 65%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Nvidia reported $96.2 billion in quarterly revenue, up 106% from a year ago, and data center revenue more than doubled to $89 billion, growing 117%.
- [2]
Management expects another $108 billion in revenue next quarter and has told investors to expect roughly 70% growth in the next fiscal year; the outlook for next year is well ahead of what Wall Street had been expecting.
- [3]
Nvidia had committed $279 billion to future supply and capacity at the end of July; three months earlier the figure was $119 billion, and at the end of January it reported $95.2 billion of manufacturing, supply and capacity commitments.
- [4]
Osman writes that nothing in the latest quarter suggests customers are delaying or slowing adoption, and that demand still looks exceptional.
- [5]
The commitments stretch several years out, with roughly $92 billion scheduled for the remainder of this fiscal year, $87 billion for fiscal 2028 and another $88 billion for fiscal 2029.
- [6]
The commitments are reportedly for memory and manufacturing capacity needed for current and future generations of Nvidia's data center systems, and some can be cancelled, rescheduled or adjusted; Osman writes it would be wrong to describe the $279 billion figure as debt.
- [7]
Osman writes that high-bandwidth memory is one of the important bottlenecks in AI infrastructure, and that securing supply competitors cannot access can itself become an advantage, letting Nvidia ship more systems, bring new architectures to customers quickly and reduce the risk that shortages create an opening for alternative technologies.
- [8]
Osman argues that by reserving capacity years in advance, Nvidia now has to be increasingly accurate about how much its customers will eventually need, rather than simply responding to AI demand as it arrives.
- [9]
Nvidia points out that customers can delay new architectures, struggle to finance infrastructure or adopt technology more gradually than expected, any of which could push revenue further out while some supply commitments remain in place.
- [10]
Nvidia's Vera Rubin platform is moving into full production, with systems running at partners including Google Cloud, Microsoft Azure, Oracle and CoreWeave, and demand is spreading beyond the hyperscalers into AI labs, enterprises, sovereign customers and industrial users.
- [11]
Data center revenue of $89 billion is 92.5% of the $96.2 billion quarterly total.
- [12]
The source does not identify which fiscal year the roughly $92 billion applies to or how many quarters remain in it, so no per-quarter figure can be computed from the material.
- [13]
Commitments rose by $160 billion in three months, an average of about $53 billion a month.
- [14]
The commitment figure is 2.34 times the level of three months earlier and about 2.9 times the level reported at the end of January.
- [15]
Semiconductor supply chains have long lead times, and semiconductor companies have always had to commit to foundries, memory, packaging and other suppliers long before the finished product reaches a customer.
- [16]
The account of the quarter and the commitment figures come from a column by Jim Osman published by Forbes.
- [17]
The three named scheduling windows account for $267 billion of the $279 billion total, leaving about $12 billion scheduled outside them.
Sources
4 independent publishers whose own reporting we read for this story.
- businessinsider.comNvidia Earnings Recap: Big Revenue Growth and 'Extreme' Memory Costs - Business Insider
1 article · August 26, 2026
- forbes.comNvidia Earnings: Q2 Revenue Beats Expectations
3 articles · August 28, 2026
- implicator.aiNvidia Forecasts 70% Growth for 2028, Cuts Margin Outlook
1 article · August 27, 2026
- newcomer.coNvidia Is Carrying the AI Economy. Is That a Problem?
1 article · August 28, 2026
Topics and entities
Topics
- AI vendor financing
- Semiconductor Supply Chain
- AI Infrastructure
- High-Bandwidth Memory