AI infrastructure supply and demand: what 42 issuers disclosed
Demand is contracted years ahead. Supply is capped by memory now and by electricity from 2027. The gap is closed by price, not by volume.
Across the last two reported quarters, the companies that build and rent AI compute describe the same condition in different words: they have sold more than they can deliver. This post screens 42 issuers across the stack, from the memory that goes into an accelerator to the electricity that runs the building it sits in, and records what each one disclosed about supply, capacity and price.
30of 42 issuers name a supply
or capacity constraint in their disclosure
5wrote that language into a
risk factor rather than an earnings call
2report no constraint anywhere
in the stack
1. What the filings say
- Demand is booked, not forecast. $2.33T of contracted future revenue sits behind $730B of guided 2026 hyperscaler capex, up 78% on 2025. OpenAI has committed about $750B of compute through 2030 and Anthropic runs at $65B, bought before the capacity exists.
- Memory is the binding constraint today. HBM, or high bandwidth memory, takes about three times the wafer capacity of standard DRAM per gigabyte, with no new capacity before 2027. AMD was first anywhere to write “industry wide memory shortage” into a risk factor. No peer used it in 2023 to 2025.
- From 2027 the constraint becomes electricity. A median 61 months runs from grid connection request to operation. The PJM auction cleared at its ceiling and still fell 6,623 MW short, and Microsoft now names power availability as a risk.
- The gap clears on price, not on volume. DRAM contract prices rose 93% to 98% in Q1 2026 and 58% to 63% in Q2. Micron’s volume grew low single digits while its gross margin reached 85%. Digital Realty renewals reprice at 25.4%.
- What ends it is financing, not demand. Neoclouds hold about $45B to $50B of liquidity against $165B to $175B of contracted build, of which 21% runs today. CoreWeave’s borrowing moved from 2.25 to 5.50 points over the benchmark in five months, with tests it must keep passing.
2. How tight each layer is
Severity is not spread evenly across the stack. It is concentrated in memory, which is fully committed, and in the hyperscalers buying from it. The layers that sell equipment into the shortage report order books rather than constraints, and the names with no constraint at all sit at the edges of the stack.
Acute: cannot serve the demand in front of itElevated: tight, and responding on price or ordersWatch: no stated constraint, or a financing item
| Layer | Issuers | Acute | Elevated | Watch or none |
|---|---|---|---|---|
| Compute silicon | 5 | 0 | 5 | 0 |
| Memory and storage | 6 | 6 | 0 | 0 |
| Foundry, equipment and packaging | 5 | 0 | 5 | 0 |
| Hyperscalers | 6 | 5 | 0 | 1 |
| Neoclouds and datacenter landlords | 7 | 2 | 1 | 4 |
| OEM, power, networking and REITs | 13 | 2 | 8 | 3 |
| All layers | 42 | 15 | 19 | 8 |
3. The mechanism
Four forces set the price of compute: what buyers are contracted to take, what the supply chain can physically deliver, what the gap between them does to price, and what would break the arrangement.
What drives demand
- Inference, not training. Huang: “Now, compute is revenue.” Volume scales with how much the models are used rather than with training runs.
- Labs contract ahead. OpenAI has committed about $750B of compute through 2030 and Anthropic runs at $65B, bought before capacity exists.
- Hyperscaler capex. $730B guided for 2026 across Microsoft, Alphabet, Amazon and Meta, up 78% on 2025, against $2.33T of forward revenue.
- Revenue per gigawatt. About $18B per GW on Hopper, $25B on Blackwell, $40B on Vera Rubin, so a megawatt earns more each time.
- A third buyer class. Sovereign and enterprise buyers sit alongside hyperscalers and neoclouds. AWS added 2 million GPUs on 27 August.
What limits supply
- Memory. HBM, high bandwidth memory, takes three times the wafer capacity of standard DRAM per gigabyte. No new capacity before 2027.
- Power. A median 61 months from grid connection request to operation. The PJM auction cleared at the ceiling and still fell 6,623 MW short.
- Substrates. Marvell paid $870M of deposits to reserve wafer and substrate capacity. Intel names glass substrates as a constraint.
- Equipment. Tools now take 12 to 24 months against about 6 before, and ASML lifts output of standard EUV machines about 30% in 2027.
- Packaging. The one link easing. The shortfall narrows from about 20% to about 10% by the end of 2026 and stops binding during 2027.
What it does to price
- Memory contracts. DRAM contract prices rose 93% to 98% in the first quarter of 2026 and 58% to 63% in the second. HBM is guided 50% up.
- Accelerators. NVIDIA told customers to expect increases above 15% on Vera Rubin and Grace Blackwell systems from the start of 2027.
- Vendor margin. NVIDIA guided gross margin to 74.0% in the third quarter and a 71 to 72% trough in the fourth. Memory is the cause.
- Rental. One year H100 contracts are up about 38% from the October 2025 low, and CoreWeave raised list prices 25% across its range in July.
- Floor space. Digital Realty renewals repriced from 5.0% to 25.4% in one quarter, against vacancy of about 1% across North America.
- Who absorbs it. Hyperscalers hold contracts that cap their increases. Dell gross margin fell 3.3 points and Supermicro has half a point left.
What would end it
- Financing. Neoclouds hold about $45B to $50B of liquidity against $165B to $175B of contracted build. 21% of it runs today.
- Credit. CoreWeave borrowing moved from 2.25 to 5.50 points over the benchmark in five months, with tests the company must keep passing.
- Working capital. NVIDIA receivables rose 55% in one quarter and operating cash flow halved to $24.1B against $59.7B of net income.
- One flat order book. Oracle leases signed but not started held at $260B. Vertiv stopped reporting backlog after the fourth quarter of 2025.
- Downstream. NVIDIA attributes weaker consumer sales to “elevated memory and systems prices”, so households carry part of the cost.
4. The company screen
42 issuers across six layers of the stack. The number that matters and the quotation are drawn from each company's last two reported quarters. Every figure is from an SEC filing unless the row says otherwise.
Compute silicon Capped by supply
| Ticker | Company | Filed | Signal | Reason for the signal | What they said |
|---|---|---|---|---|---|
| NVDA | NVIDIA | Q2 FY27, 26 Aug | Tightening | Revenue $96.2B, up 106%. Datacenter $89.0B, up 117%. Binding supply commitments $279B, from $119B a quarter earlier. Q3 guided to $108B. | The filing states: “we are currently experiencing certain supply constraints.” Kress: commitments rose “primarily related to the procurement of memory.” Gross margin now guided down, to 74.0% in Q3 and a 71 to 72% trough in Q4. FY28 revenue guided to about +70%, which Huang framed as a supply number against roughly 100% demand. |
| AVGO | Broadcom | Q2 FY26, 9 Jun | Locked supply | Binding commitments $128.1B, from $132M at year end. AI chips $10.8B, up 143%. | Tan: bookings “over $30 billion against the $10.8 billion we shipped”. Visibility runs to 2028, from 2027. |
| AMD | Advanced Micro Devices | Q2 26, 4 Aug | Tightening | Datacenter $6.72B, up 107%. Binding commitments $30.28B, from $9.4B. | The first company anywhere to write “industry wide memory shortage” into a risk factor. No such mention exists across the peer group for 2023 to 2025. Su says AMD is sharing the cost increase with customers. |
| MRVL | Marvell Technology | Q1 FY27, 28 May | Tightening | $870M of deposits to reserve wafer and substrate capacity, plus $1B of prepayments, on $2.4B quarterly revenue. | Koopmans: “Everything that touches AI has been constrained basically since the beginning of this.” The only name paying cash to reserve substrates. |
| INTC | Intel | Q2 26, 23 Jul | Tightening | Datacenter and AI $6.26B, up 59%. Total revenue up 25%, best in 15 years. | Zinsner cites “one of the most severe supply constraints in its history across leading edge logic, silicon wafers, memory, and substrates.” |
Memory and storage The bottleneck
| Ticker | Company | Filed | Signal | Reason for the signal | What they said |
|---|---|---|---|---|---|
| MU | Micron Technology | FQ3 26, 24 Jun | Sold out | Revenue $41.5B, gross margin 85%. Prices up about 60% on the quarter. $22B of customer deposits and commitments. | Mehrotra says tightness runs “beyond calendar 2027”: “we do not have line of sight as to when memory supply will be able to catch up.” |
| SNDK | SanDisk | FY26, 17 Aug | Sold out | Revenue $20.2B, up 175%. Datacenter up 437%, and per gigabyte up 150%. | Contracts with eight customers total “a minimum of $93.9 billion, assuming floor pricing”, so that is the worst case. Flash stays “on allocation beyond calendar year 2027”, meaning rationed. |
| WDC | Western Digital | FY26, 14 Aug | Sold out | Revenue up 36%: 25% more capacity, 8% higher prices. Cloud is 89%. | Tan is negotiating contracts “for calendar year 2029, 2030, and 2031”. Blended price per terabyte went from high single digits to high teens in a quarter. |
| STX | Seagate Technology | FY26, 4 Aug | Sold out | Revenue $12.20B, up 34%. Gross margin 45.6%, up 11 points. Drive capacity up 40%. | Mosley says capacity is “allocated into calendar 2028” and customers want 2029. Romano: the gap between supply and demand “is now a little bit bigger”. |
| Samsung | Samsung Electronics | Q2 26, not SEC | Sold out | Chip division operating profit KRW 89.2 trillion, a record, with the first HBM4E samples shipped to customers in the quarter. | “Proactively addressing AI demand despite limited capacity”, with a focus on servers. |
| SK hynix | SK hynix Inc. | Q2 26, not SEC | Sold out | Operating margin 76%. HBM4 shipments began. About 10 named long term customers. $26.5B raised for capacity. | CEO Kwak: “2027 will represent the worst supply shortage in the memory industry’s history”, with demand above capacity beyond 2030. |
Foundry, equipment and packaging Eases in 2027
| Ticker | Company | Filed | Signal | Reason for the signal | What they said |
|---|---|---|---|---|---|
| TSM | Taiwan Semiconductor | Q2 26, 16 Jul | Tight, expanding | Revenue $40.2B, gross margin 67.7%. 2026 capex raised to $60B to $64B, with 10% to 20% on packaging. July revenue up 44.7%. | Wei: “Our packaging capacity is so tight that now it’s limiting my customers’ growth.” Capex ahead will be “even more significantly higher”. |
| ASML | ASML Holding | Q2 26, 15 Jul | Output capped | Sales €9.3B. Full year raised to €43B to €45B. EUV output up about 30% in 2027. | Order intake “remained extremely strong”. The ceiling is ASML’s own output, not demand: it caps how fast anyone adds leading edge capacity. |
| AMAT | Applied Materials | FQ3 26, 20 Aug | Order boom | Revenue $9.12B, up 25%. Memory equipment up 52%. Next quarter up 51%. | Dickerson: “Customers continue to give us longer visibility than we’ve ever had, with some conversations extending to 2030.” Deposits are now paid up front. |
| LRCX | Lam Research | FY26, 7 Aug | Order boom | Quarterly revenue $6.72B, gross margin 52%, the best in 20 years. Next quarter guided to $8.1B, up more than 20%. | Bettinger puts 2026 industry equipment spending “in the low $150 billion range, up from our prior outlook of $140 billion”. Archer sees “an extraordinary setup”. |
| KLAC | KLA Corporation | FY26, 6 Aug | Order boom | Full year revenue $13.58B. Next quarter $4.0B. Packaging revenue up 70%. | Raised its 2026 equipment estimate to about $150B, matching Lam. Points to “rising complexity and performance specifications in memory”. |
Hyperscalers Cannot meet demand
| Ticker | Company | Filed | Signal | Reason for the signal | What they said |
|---|---|---|---|---|---|
| MSFT | Microsoft | FY26, 29 Jul | Short of capacity | Capex $140.6B including leases, up 65%. Leases signed but not started $329.1B, from $92.7B. Contracted future revenue $678B. | Hood: “Customer demand continues to exceed available capacity.” Microsoft added a paragraph on power availability found in no earlier filing. Extending building life to 25 years cuts capex by $15B. |
| GOOGL | Alphabet | Q2 26, 22 Jul | Short of capacity | 2026 capex $195B to $205B, raised twice this year. Contracted future revenue $519.5B, of which $513.9B is Cloud. | Pichai: “We continue to be supply constrained.” Google is renting third party capacity as a stopgap. It also began selling its own TPU chips into customers’ datacenters, making it a chip vendor for the first time. |
| AMZN | Amazon | Q2 26, 30 Jul | Short of capacity | Trailing capex $173.0B, up 61%. 2026 guidance raised to about $220B from about $200B. Contracted future revenue $496B. | Jassy on the raise: “The higher cost of memory pushing this number up.” On capacity: “and I believe this dynamic will also be true in 2027, too.” |
| META | Meta Platforms | Q2 26, 29 Jul | Short of capacity | 2026 capex $130B to $145B. Leases signed but not started $279.0B, plus about $68B more in July. Binding commitments $349.3B. | Zuckerberg: “there’s just nowhere near enough compute for all the demand.” Li says Meta is constrained “for the foreseeable future”. |
| ORCL | Oracle | FY26, 22 Jun | Slow conversion | Contracted future revenue $638B, from $138B. Capex $55.7B, up 162%. Free cash flow negative $23.7B. GPU utilisation 97.5%. | Magouyrk says demand exceeds supply. But only 12% of that $638B becomes revenue within a year, and signed leases were flat at $260B. |
| TSLA | Tesla | Q2 26, 22 Jul | Short of chips | Onsite AI training compute above 205 MW, more than doubled in the first half. | Musk: “We simply won’t have enough AI chips. So it’s crucial to solve that, we’ll have to solve memory, logic and packaging.” A buyer naming memory first. |
Neoclouds and datacenter landlords Limited by power
| Ticker | Company | Filed | Signal | Reason for the signal | What they said |
|---|---|---|---|---|---|
| CRWV | CoreWeave | Q2 26, 11 Aug | Costlier debt | Contracted future revenue $103.7B. 1.5 GW running against 4.2 GW contracted. Three customers are 72% of revenue. | Intrator: “pricing for prior generation SKUs is at or above where it was years ago.” It signed a 2020 chip into 2029 and raised list prices 25%. But its borrowing spread more than doubled in five months. |
| NBIS | Nebius Group | Q2 26, 12 Aug | Pricing power | Revenue $582M, up 454%. Run rate $3.0B. Contracts at $20M to $25M per MW a year. | Volozh: “We could sell our entire 2027 capacity on these terms today. We are deliberately not doing so.” It also extended assumed server life to 5 years. |
| IREN | IREN | Q3 FY26, 7 May | Fully contracted | Microsoft contract $9.7B for 200 MW over 5 years, 20% paid up front. About 150,000 GPUs installed or on order. | Draper: “there just are not idle GPUs that are not being used in this market.” Older chips are “all effectively fully utilized”, with pricing climbing. |
| APLD | Applied Digital | FY26, 29 Jul | Tenant risk | Backlog about $36B across 1,410 MW, against 175 MW running today. | It exited GPU rental to become a landlord. In March CoreWeave moved its leases into a ring fenced subsidiary, backed by a $50M letter of credit. |
| WULF | TeraWulf | Q2 26, 5 Aug | Limited by power | Anthropic lease 401 MW for 20 years, about $19B. Build cost $8M to $10M per MW. | Prager: “access to power becomes the defining constraint on AI infrastructure.” Note the pricing. Unrated Anthropic pays about $2.37M per megawatt a year. Amazon pays about $1.22M. |
| HUT | Hut 8 | Q2 26, 4 Aug | Cheap to fund | Contract value about $26.6B across 949 MW, one high grade tenant. $7.5B of project financing closed at investment grade. | Same asset as CoreWeave, investment grade tenant, cheaper funds. The rate gap tracks tenant credit. |
| CIFR | Cipher Mining | Q2 26, 4 Aug | Cheap to fund | Backlog $11.4B, about $7.5B of it Amazon. Black Pearl is 300 MW at about $5.5B over 15 years. Notes priced at 6.00%. | Amazon backed capacity earns less rent and funds cheaper. Bond and rent spreads agree. |
OEM, power, networking and REITs Cost squeeze
| Ticker | Company | Filed | Signal | Reason for the signal | What they said |
|---|---|---|---|---|---|
| DELL | Dell Technologies | Q1 FY27, 9 Jun | Cost squeeze | AI order backlog $51.3B, a record. AI server revenue $16.1B, up 757%. Gross margin down 3.3 points to 17.8%. | The filing blames “current limitations in capacity from memory manufacturers”. Clarke says memory spot prices are “up nearly 5.5 times” in six months and “we’re repricing, it feels like every day”. |
| SMCI | Super Micro Computer | FY26, 11 Aug | Most exposed | Quarterly revenue $11.1B, up 91%. Over $60B of new orders. Next quarter gross margin guided back to 10.4% to 10.8%. | Liang says revenue came in light “due to some short term customer delays in power shortage, cooling, and networking”. Power, not chips. |
| HPE | HP Enterprise | Q2 FY26, 1 Jun | Passing on cost | Server revenue up 32.7%, “predominantly” on price. AI orders $16.4B. | The filing names “the worldwide shortage in memory components”. Neri: “That cost environment and pricing environment will continue to be very elevated in 2027”, with upside down to supply. |
| ANET | Arista Networks | Q2 26, 4 Aug | Stocked up | Revenue $3.04B, up 37.7%. Multiyear purchase commitments went from $3.6B to $9.7B. Gross margin down 2.2 points. | Ullal: “The industry is going to have a 2 year problem. I don’t think we get out of it as an industry until 2028.” Arista prepaid for supply using its balance sheet. The cost is 2.2 points of margin. |
| VRT | Vertiv Holdings | Q2 26, 29 Jul | Disclosure gap | Revenue $3.27B, up 24%. Adjusted operating margin 22.6%, up 4.1 points. | It has stopped disclosing backlog, orders and book to bill since Q4 2025, when they were $15.0B, up 252%, and 2.9 times. No resumption date given. |
| ETN | Eaton | Q2 26, 30 Jul | Order boom | Americas electrical orders up 41% on a rolling year. Global backlog up 103%. | A pipeline of about 307 GW of US datacenter negotiations, roughly 15 years of construction at 2025 build rates, up from 12 years three months earlier. The 307 GW figure is second hand. |
| GEV | GE Vernova | Q2 26, 22 Jul | Sold out to 2030 | Orders $24.2B, up 88%. Turbines under contract 116 GW, guided to 125 GW plus. | Strazik: “mostly sold out through 30”, already selling 2031 slots. First half orders priced 20% above Q4 2025. Only 20% is datacenters. |
| EQIX | Equinix | Q2 26, 29 Jul | Pricing power | Gross bookings up 23%, second highest ever. Revenue per cabinet $2,538, up 6%. Capex outlook raised to $5B to $7B a year. | Equinix controls 3 GW of land where power is contracted or nearly so. It pulled more than 7,000 cabinets forward from 2027 into 2026. Space sold but not yet installed is at a record level. |
| DLR | Digital Realty | Q2 26, 23 Jul | Pricing power | Uplift on renewing leases went from 5.0% to 25.4% in one quarter. Small leases went from $247 to $280 per kilowatt per month. | This is what expiring leases reprice to in a market with 1% vacancy where grid power is booked through 2030, the only new supply reaching it. |
| COHR | Coherent | FQ4 26, 12 Aug | Output capped | Revenue $2.05B, a record, up 42%. Datacenter up 66%, and 79% of total. | Anderson: indium phosphide capacity is “our primary constraint”, the material in optics lasers. |
| ALAB | Astera Labs | Q2 26, 4 Aug | No constraint | Revenue $392M, up 104%. Next quarter guided to $540M to $560M, up 40%. PCIe 6.0 is more than half of revenue. | Content per accelerator “could grow well beyond $1,000”, against $50 to $100 at the IPO. Reports no supply constraint of its own. |
| CRDO | Credo Technology | FY26, 15 Jun | No constraint | Revenue above $1.3B, more than tripled. Raw materials inventory up 405%. Next year’s gross margin guided flat. | No stated constraint, no cost inflation, no margin compression. Stockpiling inputs ahead of “a very aggressive ramp”. Nothing in the stack. |
| FIX | Comfort Systems USA | Q2 26, 23 Jul | Build boom | Backlog $14.06B, from $8.12B like for like. Technology is 58% of first half revenue, up from 40% a year earlier. | Factory capacity expanding to about 5 million square feet by late 2027, explicitly for “current customers and existing orders rather than speculative demand”. |
5. How to read the signals
A signal summarises what a company's own disclosure implies about how tight conditions are for that company. It is not a rating, a recommendation, or a view on price.
| Signal | Severity | Meaning |
|---|---|---|
| Capped by supply | Acute | cannot make or buy enough |
| The bottleneck | Acute | the tightest link in the chain |
| Sold out | Acute | output already committed |
| Short of capacity | Acute | demand exceeds what it can serve |
| Limited by power | Acute | held back by electricity, not chips |
| Output capped | Elevated | limited by its own factory or a scarce material |
| Eases in 2027 | Elevated | tight now, loosening on a known date |
| Cost squeeze | Elevated | paying more for inputs than it can recover |
| No constraint | Watch | nothing in its disclosure is short |
6. Method and sources
The screen covers 42 issuers across the AI infrastructure stack, using the last two reported quarters for each name, to 27 August 2026. Evidence is drawn from SEC filings and the earnings calls filed alongside them, retrieved through SEC-API.io. Samsung Electronics and SK hynix do not file with the SEC; both are included so the memory layer is complete, and their rows are marked accordingly.
Two counts in this post distinguish where a company said something. Thirty issuers name a supply or capacity constraint somewhere in their disclosure. Five put that language into the filing itself rather than an earnings call: NVIDIA, AMD, Microsoft, Dell and HPE. That difference matters, because risk factor language survives legal review and a spoken answer does not.
Source: SEC Filings via SEC-API.io
