Nvidia’s $279 Billion Bet on AI: Suppliers Plan Up to 15 Times More of Some Parts
NVIDIA has put $279 billion of commitments to suppliers behind its bet that AI demand keeps growing, $160 billion of it in just three months, its own filing shows. On October 21 it lined up suppliers to make up to 15 times as much of some parts, led by SPIL’s target for optical engines.
The short answer. NVIDIA’s agreements to purchase parts and book factory space more than doubled in one quarter, and on October 21 six suppliers announced expansion plans with NVIDIA. The same morning it showed the rest of the build-out: data centers it says install in weeks and need almost no cooling water, and chips for robotaxis, robots and phone networks. NVIDIA’s own filing also warns it may be unable to cut these commitments if demand differs from its plans.
Three numbers behind NVIDIA’s bet on AI demand
On this page
- How much has Nvidia committed to its suppliers?
- How much more are Nvidia’s suppliers getting ready to make?
- How fast can an AI data center be built?
- Do AI data centers use a lot of water?
- Do AI data centers use a lot of electricity?
- Where will Nvidia’s AI chips go next?
- What did Nvidia announce in Berlin?
- How we counted
- Frequently asked questions
How much has Nvidia committed to its suppliers?
$279 billion at the end of July, by NVIDIA’s own filing, which says the commitments are there “to meet future demand.” That was $160 billion more than three months earlier, so NVIDIA’s bet more than doubled in a single quarter.
NVIDIA’s quarterly report for its second quarter of fiscal 2027, which ended July 26, 2026, says it raised its supply and capacity commitments “from $119 billion last quarter to $279 billion as of July 26, 2026 to meet future demand” (NVIDIA Form 10-Q). These are agreements to purchase parts and book factory space, which the filing says are “primarily memory and manufacturing facilities.”
Show the data
| Item | Value | Times the smallest |
|---|---|---|
| All commitments, July 26, 2026 | $279.0B | 254x |
| Added in that quarter | $160.0B | 145x |
| KYEC advance, October 21 | $1.1B | 1.0x |
The October 21 deals illustrate the kind of capacity NVIDIA is reserving, and the one priced deal is small next to the total. By our calculation, the $1.1 billion KYEC advance is about 0.4% of the $279 billion. IBIDEN gets a “multibillion-dollar” advance, under an agreement IBIDEN announced earlier this year, with no exact amount given. The filing does not say whether these advances are inside the total.
A bet this size cuts both ways. The commitments are not all fixed. The filing says some of the agreements can be cancelled, rescheduled or changed before firm orders are placed, sometimes at extra cost. It also warns that if demand differs from its plans, “we may be unable to reduce these commitments.”
They sit beside fast-growing sales. NVIDIA’s revenue was $96.2 billion in that same quarter, up 106% from a year earlier (NVIDIA results). StockTitan has also covered the chipmakers further up the same chain (see ASML on demand for chip machines and what TSMC and Samsung’s chip numbers say).
Show the data
| Category | Revenue | Note |
|---|---|---|
| Q1 FY2026 | $44.1B | |
| Q2 FY2026 | $46.7B | NVIDIA’s Paris event, June 2025 |
| Q3 FY2026 | $57.0B | |
| Q4 FY2026 | $68.1B | |
| Q1 FY2027 | $81.6B | DSX data center design, March 2026 |
| Q2 FY2027 | $96.2B |
How much more are Nvidia’s suppliers getting ready to make?
Several times more of the parts that go into NVIDIA’s AI systems, by their own targets. The biggest is SPIL’s aim of 15 times its optical engine capacity by mid-2027, and the smallest target is double.
Show the data
| Category | Value |
|---|---|
| SPIL optics, mid-2027 | 15.0x |
| Chip connectors (Amkor), 2028 | 4.0x |
| Chip boards, 3 years | 2.5x |
| Chip testing, 2027 | 2.0x |
The supply chain release names six suppliers, each expanding a different step of making an AI chip (NVIDIA release, October 21). SPIL aims for 15 times its capacity for optical engines, parts that move data with light, by mid-2027. It hopes to double that again by the end of 2028.
Amkor, which packages chips, targets a more than 300% rise by the end of 2028 in the parts that connect chips side by side inside one package, called interposers. That is more than four times today’s level. The boards chips sit on, called substrates, are set to grow 2.5 times over three years through IBIDEN, Samsung Electro-Mechanics and Unimicron.
Testing, the step where finished chips are checked, is set to double from 2026 to 2027. NVIDIA is paying KYEC $1.1 billion in advance to hold testing capacity at new U.S. facilities, with high-volume production targeted for April 2027. “Demand is extraordinary. Capacity is the constraint,” said Jensen Huang, NVIDIA’s chief executive.
These are the suppliers’ and NVIDIA’s stated targets, not our forecast. A multiple of capacity is not a multiple of chips either. Each target measures a different step on its own clock. Each new generation also packs more into a single part, and the release gives no starting capacity. So none of these converts into a number of chips or dollars of sales.
Show the data
| Date | Event |
|---|---|
| Early 2027 | First factory-built DSX POD data centers delivered |
| Feb 2027 | Expanded chip testing in Singapore at high volume |
| Apr 2027 | KYEC’s new U.S. chip testing at high volume |
| Mid-2027 | SPIL optical engine capacity 15 times today’s |
| Q1 2028 | IBIDEN’s Ono plant starts adding substrate capacity |
| Mid-2028 | Unimicron substrate expansion at full capacity |
| End of 2028 | Amkor interposer wafer capacity more than 4 times today’s |
How fast can an AI data center be built?
Another part of the build-out is faster data centers. NVIDIA says its factory-built DSX POD could be installed in as little as 14 days, against about six months today, once a site is ready.
A DSX POD is a complete AI data center packed into modular containers. Partners such as Dell Technologies, HPE, Lenovo, Schneider Electric and Vertiv do the cabling, piping and testing in their factories. The units then “arrive ready to connect” to a site’s power and cooling (NVIDIA release, October 21). Deliveries are planned to start in early 2027.
Six months against 14 days is about 13 times faster, by our arithmetic, and it is NVIDIA’s best case, since the release says “as little as.” The release lists partners that will build and deliver the units, but no price for a POD.
NVIDIA ties the speed to money. Huang said a data center “creates value when it starts producing tokens,” the units of text and data an AI model turns out. In the release’s logic, every week saved on installation is a week of earning sooner.
The faster install does not touch the slowest steps. The 14 days start at “a prepared site,” with power, cooling and networks already in place, so finding land, getting permits and connecting to the grid are not part of that count.
Do AI data centers use a lot of water?
Sites cooled by towers lose water all the time, because the towers work by evaporating it. NVIDIA says its new design could save 54 million liters a day at a 2-gigawatt site compared with cooling towers.
A cooling tower takes heat out of a building by letting water evaporate, the way sweat cools skin, and the lost water has to be replaced. NVIDIA’s design swaps that for a sealed loop. Cooling liquid flows over the chips and keeps circulating, and outdoor dry coolers, which NVIDIA describes as “similar to large radiators,” release the heat into the air “without evaporating water” (NVIDIA release, October 21).
- Liquid takes the heat Cooling liquid flows past the computer chips and absorbs their heat, at up to 45 degrees Celsius.
- A sealed loop The same liquid keeps circulating in a closed loop, recirculated instead of evaporated.
- Radiators release the heat Outdoor dry coolers, which work like large radiators, push the heat into the air without evaporating water.
- The hottest days Local climate decides whether extra mechanical cooling is needed during the hottest periods.
NVIDIA says its Vera Rubin systems can be cooled entirely by liquid at up to 45 degrees Celsius (113 degrees Fahrenheit). A radiator only sheds heat into air cooler than the liquid inside it, so the warmer the liquid, the more often outside air can take the heat. The release is open about the limit: “Local climate determines whether supplemental mechanical cooling is needed during the hottest periods.”
The releases give no overall figure for how much water AI data centers use. NVIDIA pictures its saving as the daily water use of 180,000 households, which works out to 300 liters per household. By our reading, that is close to a European home of two or three people. The European Environment Agency cites about 124 liters a day per resident, from EurEau figures (EEA).
An American reader can picture it a second way. The EPA says the average American family uses more than 300 gallons of water a day at home (EPA WaterSense), about 1,136 liters. At that rate the same 54 million liters would cover the home use of about 48,000 American families, by our count, because American homes use more water than European ones.
Show the data
| Category | Value |
|---|---|
| NVIDIA’s households | 180K |
| U.S. families (EPA) | 48K |
The claim covers the cooling water a site evaporates. It gives no figure for the water that fills and tops up the loop. Nor does it count the water power plants use to make electricity, or the water used to make the chips.
Do AI data centers use a lot of electricity?
They use a small share of the world’s power, and it is growing fast. The International Energy Agency counted data centers of every kind at about 1.5% of the world’s electricity in 2024 and expects their use to more than double by 2030.
The IEA’s 2025 report on energy and AI covers all data centers, AI and otherwise. It puts their use at 415 terawatt hours in 2024 and about 945 by 2030 (IEA, Energy and AI). The same report warns that grids are already under strain, and that about 20% of planned data center projects could face delays unless that is addressed.
NVIDIA’s answer this morning was software. It says a tool called DSX MaxLPS, which shifts power between machines inside a data center, can produce the same AI output with up to 30% less power. Eight cloud operators, including CoreWeave, Google Cloud, Lambda and Nebius, are exploring it, in the release’s words.
Two of them have early results in the release, both stated as more output within a fixed power budget. Lambda reported 24% more output in a test, and Nscale reported 49% more on four racks of NVIDIA machines in Iceland. Both are the companies’ own tests on different hardware, and they measure output gains, a different yardstick from the headline’s “up to” 30% power cut.
| Test | Hardware | Result in the release |
|---|---|---|
| Lambda proof of concept | NVIDIA HGX B200 | 24% more cluster token throughput, 19% more GPU capacity, within a fixed power budget |
| Nscale evaluation, Iceland | Four GB300 NVL72 racks, 264 kilowatts | 49% higher inference throughput, 37% more managed GPU capacity, within the same provisioned power budget |
| Headline claim | DSX MaxLPS | Same number of tokens with up to 30% less power |
Source: NVIDIA release of October 21, 2026 (DSX). The tests report output gains within a power budget; they are not measured power cuts.
A second tool, DSX Flex, lets a data center cut its draw when the grid is under strain. In a trial with Britain’s National Grid and others, NVIDIA says a cluster cut its power draw by up to 40% during simulated grid events while critical work kept running.
Efficiency and growth pull in opposite directions. Suppose a data center does twice as much AI work next year and each piece of work needs 30% less electricity. It would still use 1.4 times the electricity it uses today, because 2 times 0.7 is 1.4.
Where will Nvidia’s AI chips go next?
Beyond data centers, NVIDIA is pointing its chips at robotaxis, humanoid robots, phone networks and hospitals. Most of it is trials and tools for builders, dated 2027, and a few pieces are available now.
The biggest number is on the road. Bolt plans to start gathering driving data in Madrid in 2027, subject to regulatory approval. That is part of a planned European rollout of at least 25,000 self-driving-capable Lucid vehicles (NVIDIA release). Mercedes-Benz and Volvo Cars are also building on NVIDIA’s Hyperion car computer and Alpamayo driving software, and Volkswagen’s van business showed its self-driving ID. Buzz on Hyperion (see our article on Alpamayo at CES 2026).
For robots, Agile Robots unveiled a humanoid for researchers, 1.40 meters tall and due in early 2027. NVIDIA also made its Isaac Lab 3.0 robot training software generally available (NVIDIA release). Huang said “the ChatGPT moment for robotics is fast approaching.”
In phone networks, T-Mobile made what NVIDIA calls the world’s first AI-RAN call on a commercial network, in Bellevue, Washington, with Nokia (NVIDIA release). AI-RAN means running the radio part of a mobile network on the same AI computers that can serve other work. BT plans live trials in Birmingham in 2027.
In health care, a new open model for designing proteins, built with the University of Washington’s Institute for Protein Design, is available today (NVIDIA release). And OpenShell, available now, is software and chips meant to keep AI agents, programs that take actions on their own, inside set limits. NVIDIA lists Anthropic, Microsoft and OpenAI among those working with it (NVIDIA release).
Our reading is that these five releases describe possible demand for the hardware in the other three, since each new use runs on NVIDIA computers. The releases do not say how many extra chips any of them will need, or how much more water and power.
| Announcement | Status in the release |
|---|---|
| OpenShell safe-agent design | Now available |
| RFD4-Proteina protein design model | Available today through partners |
| Isaac Lab 3.0 robot training software | Now generally available |
| DSX POD factory-built data centers | Deliveries worldwide in early 2027 |
| Agile ONE SN research humanoid | Scheduled for early 2027 |
| Bolt robotaxi data gathering in Madrid | 2027, subject to regulatory approval |
| BT live AI-RAN trials, Elisa commercial AI-RAN services | 2027 |
Source: NVIDIA releases of October 21, 2026, as worded in each release.
What did Nvidia announce in Berlin?
Eight press releases in 59 minutes, from 05:34 to 06:33 ET on October 21, during NVIDIA’s developer event in Berlin. One was about suppliers, two about building and running data centers, one about keeping AI agents in check, and four about new places for AI.
The full list, with each release on StockTitan, is in the table below. Most figures in them are NVIDIA’s or partners’ own claims, and the releases cite no independent test of the products.
The numbers in these releases also differ from last year’s. At its Paris event in June 2025, one NVIDIA release spoke of 18,000 Grace Blackwell systems for the first phase of a Mistral AI cloud (NVIDIA release, June 2025). This morning’s releases count supplier capacity, days, liters and watts instead, in the examples we compared.
| Time (ET) | Release | Main claim |
|---|---|---|
| 05:34 | DSX POD AI factories | Install time at a prepared site from six months to as little as 14 days; deliveries early 2027 |
| 05:35 | European AI factories with NVIDIA DSX | Up to 100% less cooling water; up to 30% less power for the same output |
| 05:48 | OpenShell reference architecture | Safety design for AI agents, available now |
| 06:11 | Telecom networks with Nokia | T-Mobile AI-RAN call; nine operators in Europe, the Middle East and Africa building or expanding |
| 06:13 | Healthcare and life sciences | RFD4-Proteina open protein model; Siemens Healthineers radiotherapy work |
| 06:14 | Self-driving and robotaxis on Hyperion | Bolt, Mercedes-Benz, Volvo Cars, Volkswagen Commercial Vehicles, Einride |
| 06:20 | Isaac robotics platform | Agile ONE SN humanoid; Isaac Lab 3.0 |
| 06:33 | Supply chain capacity | SPIL optical engines 15 times by mid-2027; Amkor more than 300% by end of 2028; substrates 2.5 times in three years; $1.1 billion advance to KYEC |
Source: StockTitan news pages for NVIDIA’s releases of October 21, 2026; times as published on StockTitan.
How we counted
Supplier targets: a 300% increase is four times the starting level. SPIL’s 15 times followed by another doubling would be 30 times if both are reached, and both are company targets. Commitments: $279 billion minus $119 billion is $160 billion, and $1.1 billion divided by $279 billion is 0.39%. Install time: six months is about 182 days, and 182 divided by 14 is 13.
Water per household: NVIDIA’s 54 million liters a day divided by its 180,000 households is 300 liters each. American families: 54 million liters divided by 3.785 liters per U.S. gallon is about 14.27 million gallons. Divided by the EPA’s 300 gallons per family a day, that is about 47,550 families, rounded to 48,000. Because the EPA says “more than 300 gallons,” the true number of families is somewhat lower. Release count: the eight releases are the NVIDIA items in StockTitan’s news database stamped between 05:34 and 06:33 ET on October 21, 2026.
Frequently asked questions
What is the Nvidia Amkor deal?
NVIDIA and Amkor plan to expand Amkor’s advanced chip packaging in South Korea. They target a more than 300% increase by the end of 2028 in interposer capacity, the parts that connect chips inside a package. It is one of six supplier plans in NVIDIA’s October 21 supply chain release.
How much has Nvidia committed to suppliers?
$279 billion in supply and capacity commitments as of July 26, 2026, up from $119 billion a quarter earlier, according to NVIDIA’s Form 10-Q. The filing says some agreements can be cancelled, rescheduled or changed before firm orders, sometimes at extra cost.
Do AI data centers use a lot of water?
Sites cooled by evaporating towers lose water all the time. NVIDIA says its DSX design could save 54 million liters a day at a 2-gigawatt site compared with cooling towers. It says the design, which releases heat through radiators, can bring that cooling water close to zero, with extra cooling possibly needed on the hottest days.
What is Nvidia DSX?
DSX is NVIDIA’s design for complete AI data centers, covering the computers, power, cooling and the software that runs them. On October 21 NVIDIA added DSX POD, a version built in factories inside containers, which it says can be installed at a prepared site in as little as 14 days.
Where can I read Nvidia’s October 21 press releases on StockTitan?
All eight are on StockTitan’s NVIDIA news feed, linked in the table above, alongside the company’s overview page and its financials.
Sources
- NVIDIA and Supply Chain Partners Expand Capacity to Accelerate AI Infrastructure Buildout, October 21, 2026 (StockTitan)
- NVIDIA Introduces DSX POD AI Factories to Accelerate Europe’s Infrastructure Deployment, October 21, 2026 (StockTitan)
- European Ecosystem Builds World’s Most Efficient AI Factories With NVIDIA DSX, October 21, 2026 (StockTitan)
- NVIDIA Launches OpenShell Reference Architecture for Safe AI Agents, October 21, 2026 (StockTitan)
- NVIDIA, Nokia and Leading Telecom Operators Turn Networks Into Distributed Intelligence Platforms, October 21, 2026 (StockTitan)
- Healthcare and Life Sciences Leaders Build on NVIDIA, October 21, 2026 (StockTitan)
- European Mobility Leaders Build Safe Vehicles and Robotaxi Fleets on NVIDIA Hyperion, October 21, 2026 (StockTitan)
- NVIDIA Advances Isaac Platform for General-Purpose Robotics, October 21, 2026 (StockTitan)
- NVIDIA results for the first quarter of fiscal 2026, May 28, 2025 (StockTitan)
- NVIDIA results for the second quarter of fiscal 2026, August 27, 2025 (StockTitan)
- NVIDIA results for the third quarter of fiscal 2026, November 19, 2025 (StockTitan)
- NVIDIA results for the fourth quarter of fiscal 2026, February 25, 2026 (StockTitan)
- NVIDIA results for the first quarter of fiscal 2027, May 20, 2026 (StockTitan)
- NVIDIA results for the second quarter of fiscal 2027, August 26, 2026 (StockTitan)
- Europe Builds AI Infrastructure With NVIDIA, June 11, 2025 (StockTitan)
- NVIDIA overview (StockTitan)
- NVIDIA financials (StockTitan)
- NVIDIA Form 10-Q for the quarter ended July 26, 2026 (SEC)
- International Energy Agency, Energy and AI, executive summary
- European Environment Agency, Water savings for a water-resilient Europe
- U.S. EPA WaterSense, How We Use Water
The information provided in this article is for educational and informational purposes only. It does not constitute financial advice, investment recommendation, or an endorsement of any particular investment strategy. Past performance does not guarantee future results. Investors should conduct their own research and consult with a qualified financial advisor before making investment decisions.