Manas Bihani
About

the questions

  1. What is a moat in an AI world?
  2. Why do AI products converge?
  3. What becomes scarce when intelligence becomes cheap?
  4. Does distribution matter more than technology?
  5. Why might human-made things become more valuable?
  6. What happens to expertise when everyone has the same models?
  7. Which parts of an AI startup are actually defensible?
  8. Where does value move when intelligence becomes commoditized?

everything on the desk

  1. The periodic table of the AI stackVisualization
  2. What is a moat when the model isn't yours?Note
  3. The problem-selection premiumNote
  4. Same model, different wiringNote
  5. Selection is the new bottleneckNote
  6. Get friendly with the AI raceEssay
  7. The convergence taxNote
  8. The luxury of realityNote
  9. The non-technical technical advantageNote
  10. The verification economyNote
  11. The bets against the wallVisualization
  12. You can't buy your way outVisualization
  13. How a chatbot writes one wordVisualization
  14. The grid is the last wallVisualization
  15. Who got paidVisualization
  16. Why this paper mattersExplainer
  17. Transformer: Why did transformers replace RNNs?Vaswani et al., NeurIPS 2017
  18. KV cache: Why does a long conversation get slower and cost more than a short one?Shazeer, 2019
  19. Mixture of experts: Why do some AI models have experts?Fedus, Zoph and Shazeer, 2021
  20. FlashAttention: Why is attention slow when the GPU is barely doing any arithmetic?Dao et al., NeurIPS 2022
  21. Mamba: Why does a model reread the whole conversation instead of just remembering it?Gu & Dao, 2023
  22. PagedAttention: Why does a GPU with free memory still refuse new requests?Kwon et al., SOSP 2023
  23. DeepSeek: How did DeepSeek train a frontier model so cheaply?DeepSeek-AI, 2024
  24. Jamba: Why does Jamba matter?Lieber et al., AI21 Labs, 2024
  25. BitNet: Why does BitNet matter?Ma et al., Microsoft Research, 2025
  26. DeepSeek-R1: Can a small AI model learn to reason like a huge one?DeepSeek-AI, 2025
  27. Kimi K2: Why does Kimi K2 matter?Kimi Team, Moonshot AI, 2025
  28. Sliding-window attention: How do models handle huge context windows without the memory bill exploding?Gemma Team, Google DeepMind, 2025
  29. How electricity becomes intelligenceVisualization
  30. This desk, as a datasetDataset
  31. The first version of this roomNote
  32. The aura dividendNote
  33. Distribution is rented attentionNote
  34. The Convergence TestNote
  35. A shelf for thinking about cheap intelligenceCollection
  36. Anatomy of an AI startupNote
  37. Six shocks to expertiseNote
  38. Nineteen Public KeysEssay
  39. The value migration machineModel
  40. AAA-Rated GPUsEssay
  41. Moats, before and afterVisualization
  42. The rhinoceros problemNote
  43. What becomes scarce when intelligence becomes cheap?Essay
  44. AI Has Passed Every Exam. It Has Never Had an Idea.Essay
  45. What Becomes Scarce After Intelligence?Essay
  46. India’s Carbon Markets : A New Test for Global Climate PolicyEssay
  47. Google Wants AI to Become BoringEssay
  48. The Wall That Wasn’t YoursEssay
  49. The Rate-Limiting StepEssay
  50. The Speed of Being WrongEssay
  51. Uber Burned a Year of AI Budget in Four Months. A Rat Catcher in 1902 Knew WhyEssay
  52. Finding a Flat in India Is Broken. We Have the Technology to Fix It. Nobody With Power Wants To.Essay
  53. Why We Can Never Have Good Social MediaEssay
  54. Gen Z Is Going OfflineEssay

rooms

  1. Home
  2. Writing
  3. Projects
  4. Reading & Watching
  5. All the questions
  6. Everything, as a contact sheet
  7. About

Visualization · 26 Sept 2026

Who got paid

When AI hardware ran short, some companies got to charge far more. Others sold more but earned less on each sale, and a few ended up paying more themselves. Same shortage, different endings. The difference comes down to one question: could the buyer go somewhere else?

trying to answer →Where does value move when intelligence becomes commoditized?What becomes scarce when intelligence becomes cheap?

In 2023, the year everyone was short of AI chips, NVIDIA’s gross margin jumped from 57% to 73%, and its revenue went on to grow about twentyfold from 2019. Over the same years, the companies that build those chips into racks booked tens of billions of dollars in orders. Their revenue grew too, but their margins went down. Same shortage, opposite results.

A shortage only creates the chance to charge more. Whether a company gets to take it depends on one thing: whether its buyer has anywhere else to go. Follow the money down the machine, from the chip factory to the power line, and it lands in one of a handful of ways. Each step of the machine gets two charts, and the hatched years mark when that step was the thing everyone was short of. The first chart is revenue, as a multiple of each company’s first year on file: did the shortage bring more sales? The second is gross margin, what each company keeps from every dollar of sales: did the shortage let it charge more? Most margin lines barely move, so revenue is where a shortage shows up first. The margin is what tells you who had the power to charge for it.

They charge more

Where one company holds a step that nobody can go around, the shortage turns into price.

TSMC makes most of the world’s most advanced chips. As AI chips moved onto its newest process, that capacity became something the largest companies in the world compete for, and TSMC is reported to be raising prices for 2027. It appears twice on this ladder, because it also does the step that joins the chip to its memory. Intel, the other company trying to sell cutting-edge chipmaking, watched its margin fall from 59% to 35% over the same years. Owning a chip factory is not the same as owning the one everyone needs.

NVIDIA is the clearest case, with a caution. Part of its margin is the shortage; part is its software, which makes buyers reluctant to switch even when rivals catch up. The memory makers are the purest one: three companies, years to build a new factory, and SK Hynix, which led from the start, reported keeping 76 cents of operating profit on every dollar of sales in mid-2026. Micron, the one memory maker that reports to US regulators, shows how swingy that business is: its margin fell below zero in 2023 before the AI memory boom lifted it back.

They sell more, not dearer

Where many companies can do the job, the shortage arrives as orders, not higher prices.

Dell booked $60.9bn of AI server orders in a single quarter of 2026. Supermicro’s annual revenue went from $22bn to $39bn: its revenue line is one of the steepest on the page. Both saw their margins drift down. Every AI chip has to be built into a rack, but anyone with a factory and skilled workers can build racks, so buyers could always go elsewhere.

Cooling sits in between. Vertiv, the largest listed specialist, sold far more and did raise its margin, from 28% to 37%. The smaller specialists took a different payday: they were bought by bigger companies that already equip buildings.

They rent it out, on credit

Between the chip makers and the labs sits a newer kind of company: the neocloud. CoreWeave, Nebius, Lambda and Fluidstack buy the chips, mostly with borrowed money, and rent them out on contracts years long.

CoreWeave’s revenue went from $0.2bn in 2023 to $5.1bn in 2025, and it had about $104bn more under contract by mid-2026. None of it shows as a margin, because it reports none, and its losses grew with its interest bill. This is also where the money goes in a circle. NVIDIA invests in CoreWeave, sells it the chips, and has agreed to buy up to $6.3bn of any capacity it can’t sell. It rents 18,000 of its own GPUs back from Lambda. Oracle has $664bn of contracted revenue still to deliver, most of it from OpenAI. A shortage paid for this way makes the orders real, and also puts every company in the circle at risk if any one of them stops paying.

They’re sold out years ahead

Some things are short but sold far in advance, at prices agreed before the shortage peaked. The money is real; it’s just waiting in the order book.

GE Vernova’s order book for grid equipment more than quadrupled in four years, to $35bn, and it is taking reservations for gas turbines to be delivered in 2031. Its margin has been climbing from a low base, and Eaton’s has crept up too.

They pay more

And some companies sit right next to a shortage and end up paying for it.

Arista makes the network switches that wire AI clusters together, and in 2026 demand for them ran ahead of what it could ship. But the parts it needs were scarce too, so it said it was paying more to keep them coming and expected its margin to be squeezed. Its margin ended where it started, around 64%. Standing next to a bottleneck is not the same as owning it.

They own the land, or the right to connect

At the far end of the machine, the money stops showing up in margins at all.

Out here the scarce thing is a place: land that already has a power connection, or a spot in the line to get one. Its value sits in what the land is worth and who holds the right, not in anyone’s profit margin. That’s why buyers now pay to restart old power plants and build right next to existing ones.

So who got paid?

Not the buyers. Microsoft, the one buyer here with a clean margin line, kept its margin steady; the buyers paid for the shortage in what they spent and what they signed. The money went to whoever stood on a step nobody could go around, at the moment the shortage arrived there.

That is the lens for anyone backing a company in this chain. The question isn’t whether AI demand is real. It’s where the shortage sits this year, and whether the company standing there can make its buyers pay, or whether they will go around it, order from someone else, or wait. The shortage moves every year or two, outward from the chip toward the power line. You can’t buy your way out is about why it moves; the bets against the wall is about who is betting on where it goes next.

What the market believed

Share prices are the market’s running guess at where the shortage is. A jump is a belief, not proof that anyone got paid, but read in order they walk the same path the money did.

  1. 2022-11-30 · Demand for AIOpenAI, whose largest backer is Microsoft, releases ChatGPT
  2. 2023-05-25 · The AI chipNVIDIA rises 24% in a day, adding about $184bn, on guidance of $11bn for the quarter
  3. 2023-09-06 · Advanced packaging (chip-to-chip)TSMC's chairman says the shortage is of its CoWoS packaging capacity, not of AI chips, and will last about 18 months
  4. 2024-03-13 · Memory next to the chipTrendForce reports NVIDIA's HBM3 came at first from SK Hynix alone
  5. 2024-03-18 · CoolingNVIDIA announces GB200 NVL72, a rack that ships liquid-cooled
  6. 2024-09-20 · Power plantsConstellation rises 22% on a 20-year, 835 MW deal to restart a Three Mile Island reactor for Microsoft
  7. 2024-10-30 · Assembly into racksSupermicro falls 33% in a day after its auditor resigns
  8. 2025-01-27 · Demand for AIDeepSeek: NVIDIA falls 17%, $589bn, the largest one-day loss in US market history. Vistra falls 28%; Constellation and GE Vernova more than 20%
  9. 2025-09-22 · Memory next to the chipSamsung's 12-high HBM3E reportedly clears NVIDIA's tests, 18 months after it was developed
  10. 2025-12-24 · Memory next to the chipNVIDIA agrees to pay about $20bn to license Groq's inference technology and hire its founder
  11. 2026-02-02 · Wiring between chipsMarvell completes its $3.25bn purchase of Celestial AI, a photonic interconnect startup
  12. 2026-05-14 · Memory next to the chipCerebras lists at $185 a share and opens at $350
  13. 2026-05-27 · Memory next to the chipSK Hynix passes $1 trillion in market value; Samsung and Micron cross the same line within weeks
  14. 2026-07-23 · Power plantsGE Vernova's gas turbine backlog reaches 116 GW, and it is taking reservations for 2031 delivery
  15. 2026-07-28 · Memory next to the chipSamsung falls more than 13% and SK Hynix more than 14% in Seoul; Micron more than 8%
  16. 2026-09-01 · Power plantsGoogle agrees to buy 396 MW of enhanced geothermal power from Fervo for a Utah data centre
The whole ladder all 13 steps, from demand to the grid, with every source

  1. Demand for AI

    revenue, × its first year

    1×5×20×MicrosoftMicrosoft 2.2×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%MicrosoftMicrosoft 69%2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    The buyers. A shortage anywhere below shows up in what they spend, not in their own margins.

  2. Renting out the chips (neoclouds)

    revenue, × its first year

    1×5×20×CoreWeaveCoreWeave 22×OracleOracle 1.7×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%none of them reports a gross margin2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    The neoclouds buy chips, mostly with borrowed money, and rent them to the AI labs and the big clouds. The shortage reaches them as orders years long and as debt, not as a margin: neither company here reports a gross-profit line.

    The order book. CoreWeave had about $104bn of contracted revenue at the end of June 2026, including Meta ($14bn, then $21bn more through 2032) and OpenAI (up to $22.4bn). Its quarterly revenue doubled, and its net loss widened to $626m as interest costs grew.

    The circle. NVIDIA invests in CoreWeave, sells it the chips, and agreed in 2023 to buy up to $6.3bn of whatever capacity it cannot sell, through 2032. It also rents back 18,000 of its own GPUs from Lambda for $1.5bn over four years. The chip maker is investor, supplier and buyer of last resort at once.

    The others. Nebius signed Microsoft for $17.4bn to $19.4bn and Meta for about $3bn, then more. Oracle’s contracted but unbooked revenue reached $664bn at the end of August 2026, most of it from OpenAI.

  3. Chip factories

    revenue, × its first year

    1×5×20×short here2019TSMCTSMC 2.7×IntelIntel 0.7×ASMLASML 2.8×Applied MaterialsAMAT 1.9×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%short here46%TSMCTSMC 56%IntelIntel 35%ASMLASML 53%Applied MaterialsAMAT 49%2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    Before the package there is the wafer. TrendForce: 3nm capacity, dominated by TSMC, has become “a scarce resource fiercely contested by global tech giants”.

    The contrast. Intel, the other company selling leading-edge wafers, saw its gross margin fall from 59% to 35% over the same years. Owning a fab is not the same as owning the one everyone needs.

    The proof. TSMC is reported to be raising advanced-node prices by up to 10% for 2027, with a further premium for customers who want extra HPC capacity.

    Short here from 2025-10. AI chips moved from 4 nm to 3 nm between late 2025 and 2026, concentrating demand on the node TSMC leads.

  4. The AI chip

    revenue, × its first year

    1×5×20×short here2019NVIDIANVIDIA 20×AMDAMD 5.2×BroadcomBroadcom 2.8×MarvellMarvell 3.0×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%short here62%NVIDIANVIDIA 71%AMDAMD 50%BroadcomBroadcom 68%MarvellMarvell 51%2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    NVIDIA. Moving to another accelerator means rewriting and re-validating software, so buyers pay its price rather than wait for a substitute.

    Who sells it. One company sells the accelerator most AI software is written for. AMD, and custom chips that Broadcom and Marvell design for single buyers, are the alternatives.

    Why no one adds more quickly. NVIDIA owns no factories. How many it can ship is set by TSMC's packaging lines and the memory makers' HBM, and both take years to expand.

    The proof. Fiscal 2026 revenue of $215.9bn, about 90% of it data centre, at a gross margin near 75%.

    Not all of it is scarcity. A gross margin also carries software lock-in and a design lead. The shortage made room to charge; the software made it hard for buyers to go around.

    Short here from 2023-05 to 2023-09. The market read the shortage as a shortage of accelerators and paid the company that sells them. Within four months the foundry said the short step was packaging.

  5. Memory next to the chip

    revenue, × its first year

    1×5×20×short here2019MicronMicron 1.6×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%short here46%MicronMicron 40%2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    The three makers with qualified lines already running. Because HBM took wafers from ordinary DRAM, the shortage spread to all memory and priced the whole business up. SK Hynix and Samsung do not file with the SEC, so Micron stands in on the chart.

    Who sells it. Three companies make DRAM. SK Hynix was first to have HBM3 qualified by NVIDIA, for the H100, in June 2022, and held a large early lead; Samsung took until September 2025 to clear the tests for its 12-high HBM3E.

    Why no one adds more quickly. New HBM capacity is measured in years: SK Hynix's new Yongin Y2 cleanroom is targeted for June 2029, and its Indiana HBM plant for volume production in the third quarter of 2029. Meanwhile wafers turned into HBM come out of ordinary DRAM's share.

    The proof. SK Hynix reported a 76% operating margin in Q2 2026, on revenue up 257% in a year.

    Short here from 2023-06. SK Hynix was first to qualify HBM3 with NVIDIA and held a large early lead. By 2026 turning wafers into HBM had made ordinary DRAM scarce as well, and all three makers were valued above $1 trillion.

  6. Advanced packaging (chip-to-chip)

    revenue, × its first year

    1×5×20×short here2019TSMCTSMC 2.7×AmkorAmkor 1.7×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%short here46%TSMCTSMC 56%AmkorAmkor 14%2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    TSMC, not the packaging industry. The scarce step was TSMC's own, which is why the outsourced packagers' margins did not rise with it.

    Who sells it. The scarce thing is qualified advanced-packaging capacity at the right technology and yield. TSMC makes most CoWoS, the step that joins an accelerator to its HBM; outsourced packagers such as ASE and Amkor do the steps around it and have been adding capacity of their own.

    Why no one adds more quickly. Capacity roughly doubled each year, from about 70,000 wafers a month in 2025 toward 130,000–140,000 by the end of 2026, and demand still ran an estimated 20% ahead.

    The proof. TSMC is reported to be raising advanced-node prices by up to 10% for 2027, with a further premium for customers who want extra HPC capacity.

    Where it is going. TrendForce expects the gap between CoWoS supply and demand to narrow from about 20% to about 10% by the end of 2026, and the shortage of 2.5D packaging to begin easing in 2027. The constraint moves on; it is not removed.

    Short here from 2023-09. TSMC's own chairman named CoWoS capacity as the limit in September 2023. Capacity then grew several times over and was still fully allocated in 2026.

  7. Wiring between chips

    revenue, × its first year

    1×5×20×short here2019AristaArista 3.7×BroadcomBroadcom 2.8×CredoCredo 25×LumentumLumentum 1.8×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%short here64%AristaArista 64%BroadcomBroadcom 68%CredoCredo 68%LumentumLumentum 42%2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    Arista sells the switches every AI cluster is wired with, and demand ran ahead of what it could ship. But the scarce parts sat upstream of it: on its first-quarter 2026 call it said it was paying more “to assure supply continuity” and expected margin pressure. Its gross margin held near 63%. Standing next to a bottleneck is not the same as holding it.

    Optics. Whoever holds qualified module and laser capacity while the clusters keep growing.

    The proof. The market for AI cluster optics was estimated at $16.5bn in 2025 and $26bn in 2026; Lumentum’s revenue grew 90% in a year to its fiscal Q3 2026.

    Short here from 2025-06. Moving data between packages became expensive enough in power that a chip company paid $3.25bn for a photonic interconnect startup, and the optics makers at least doubled in the first months of 2026.

  8. Assembly into racks

    revenue, × its first year

    1×5×20×SupermicroSupermicro 6.3×DellDell 1.3×CelesticaCelestica 1.7×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%SupermicroSupermicro 11%DellDell 20%CelesticaCelestica 12%2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    The assemblers, in revenue rather than in margin. Every accelerator has to be built into a rack, so they made billions; but because many can do the work, the gap reached them as volume, not as a higher price per rack.

    Who sells it. Foxconn, Quanta, Wistron and its subsidiary Wiwynn, Supermicro, Dell, Celestica and others. Assembly is skilled work, but many can do it.

    Why no one adds more quickly. Quarters. A new assembly line is a building and a workforce, not a fab.

    The proof. Dell booked $60.9bn of AI server orders in one quarter of 2026 and held a $95bn backlog. Supermicro’s revenue reached $39.1bn in the year to June 2026, up from $22.0bn. Wiwynn’s 2025 revenue was NT$950.7bn, up 164%. Foxconn, the largest assembler, still reported a 6.15% gross margin for Q2 2026.

  9. Cooling

    revenue, × its first year

    1×5×20×short here2019VertivVertiv 2.3×nVentnVent 1.8×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%short here33%VertivVertiv 36%nVentnVent 38%2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    The largest qualified vendors, through volume and backlog more than through price.

    Who sells it. Many companies make cold plates and coolant distribution units; fewer are qualified for the racks NVIDIA specifies.

    Why no one adds more quickly. Manufacturing lines are added in months to a year or two, far faster than a fab.

    The proof. Vertiv’s order backlog more than doubled to over $15bn, and its Q1 2026 adjusted earnings rose 83%.

    Then the incumbents bought in. Flex bought JetCool (Nov 2024, undisclosed); Schneider Electric bought Motivair (75%) (Feb 2025, controlling stake); Trane bought LiquidStack (Mar 2026, undisclosed); Eaton bought Boyd Thermal (Mar 2026, $9.5bn); Ecolab bought CoolIT Systems (2026, about $4.75bn).

    Short here from 2024-03. NVIDIA announced a rack that ships liquid-cooled, at around 120 kW. Air stopped being an option for the part everyone was buying.

  10. Transformers and power gear

    revenue, × its first year

    1×5×20×EatonEaton 1.3×Monolithic PowerMPWR 4.4×FlexFlex 1.1×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%EatonEaton 38%Monolithic PowerMPWR 55%FlexFlex 9%2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    So far the makers hold it as backlog more than as margin: much of what ships was priced before the shortage peaked, which delays the rent rather than removing it.

    Who sells it. A small number of makers of large power transformers and high-voltage switchgear, all short of the same grain-oriented electrical steel.

    Why no one adds more quickly. Power transformers averaged 128 weeks to deliver and generator step-up transformers 144 weeks in 2025; substation transformers passed 160 weeks in 2026.

    The proof. US transformer lead times reached as long as four years by 2026.

    The order book. GE Vernova’s electrification equipment backlog more than quadrupled in four years, to $35bn.

  11. Buildings and land

    revenue, × its first year

    1×5×20×the value is in land, not in a margin2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%the value is in land, not in a margin2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    Owners of sites that already have power, which is why buyers now pay to restart plants and to build beside existing ones.

    Who sells it. Anyone with land and capital can build a hall. The scarce thing is a site that already has a power connection.

    Why no one adds more quickly. A building takes a year or two; the connection it needs takes several.

    The proof. Microsoft contracted for a restarted Three Mile Island reactor, and Meta agreed to prepay Oklo for power at a site Oklo already owned in Ohio.

    Why there is no line. The rent here is in what a powered site is worth, an asset price rather than a gross margin, so a margin chart cannot show it.

  12. Power plants

    revenue, × its first year

    1×5×20×short here2022GE VernovaGE Vernova 1.3×2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%short here12%GE VernovaGE Vernova 20%2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    Whoever already owns generation that runs around the clock, and the turbine makers through years of reserved slots.

    Who sells it. Three companies make heavy-duty gas turbines: GE Vernova, Siemens Energy and Mitsubishi Heavy. A handful own the nuclear plants already running.

    Why no one adds more quickly. A heavy-duty gas turbine ordered in 2026 is delivered around 2031. A new reactor takes longer.

    The proof. GE Vernova's gas backlog and slot reservations reached 116 GW in Q2 2026, and it is taking reservations for 2031.

    Short here from 2024-09. Buyers began paying to restart shut reactors and reserving turbines years ahead rather than wait in the connection queue. A heavy-duty gas turbine ordered in 2026 arrives around 2031.

  13. Plugging into the grid

    revenue, × its first year

    1×5×20×short herea right to connect has no margin2019’192020’202021’212022’222023’232024’242025’25

    gross margin

    0%40%80%short herea right to connect has no margin2019’192020’202021’212022’222023’232024’242025’25
    the evidence

    A place on the grid is a right, not a product. It belongs to whoever already holds a connection or a place in the queue, which is why buyers now pay for sites and restarted plants that come with one.

    The queue. At the end of 2025 about 8,200 projects, 1,312 GW of generation and about 749 GW of storage, were waiting to connect to the US transmission grid.

    A caveat. Those are queues for power plants, not for data centres, which wait in their utility’s own process. They show how long the grid takes to add supply, not how long any one data centre waits.

    Short here from 2024-09. Buyers began paying to restart shut reactors and reserving turbines years ahead rather than wait in the connection queue. A heavy-duty gas turbine ordered in 2026 arrives around 2031.