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

The grid is the last wall

Inside the fence, everything a data centre needs can be bought, eventually. Outside it, the power has to come from somewhere, and every way of getting more is measured in years. An engraved walk from the data hall out along the line to the gas turbines, the reopened reactor and the geothermal wells that buyers are now paying for, then country by country: where the wall is highest.

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

The constraint in AI hardware has been walking outward for three years: from the chip to its memory, to the package, to the rack, to the building. This is where it stops walking, because past the fence nothing can be ordered on a timescale that matters to a model. Scroll, and follow the line out.

  1. inside the fence

    Everything inside the fence can be bought

    So a new data centre starts from one number, how much power can I get, and everything else is drawn around it. Two to four years from land to first power, at roughly ten million dollars a megawatt.

    IEA, Energy and AI (2025)

  2. the substation

    The transformer is the first thing money cannot hurry

    Power transformers averaged 128 weeks to deliver and generator step-up transformers 144 weeks in 2025; substation transformers passed 160 weeks in 2026. US transformer lead times reached as long as four years by 2026.

    POWER Magazine, Transformers in 2026: shortage, scramble, or self-inflicted crisis? · pv magazine USA, transformer lead times extend to four years (11 May 2026)

  3. the line

    Then the queue to connect at all

    Nobody queued for interconnection when a large site was 30 MW and utilities had spare capacity. At gigawatt scale the interconnection study decides the schedule, and it runs in years while model generations run in months.

    LBNL, Queued Up — interconnection queues · IEA, Energy and AI (2025)

  4. 2026-07-23

    Gas: a turbine ordered now arrives around 2031

    GE Vernova's gas turbine backlog reaches 116 GW, and it is taking reservations for 2031 delivery. A turbine ordered now arrives in about five years. Anything with a shorter lead time gets solved first, which is why the constraint ends up here.

    Utility Dive, GE Vernova gas turbine backlog climbs to 116 GW (2026)

  5. 2024-09-20

    Nuclear: reopen what was shut

    Constellation rises 22% on a 20-year, 835 MW deal to restart a Three Mile Island reactor for Microsoft. A data centre buyer pays to reopen a shut nuclear unit rather than wait in the connection queue. Power becomes something bought years ahead.

    CNBC, Constellation Energy to restart Three Mile Island (20 Sep 2024)

  6. 2026-09-01

    Geothermal: drill for a new kind of plant

    Google agrees to buy 396 MW of enhanced geothermal power from Fervo for a Utah data centre. At the end of the path the constraint is still the connection, and buyers now pay for new kinds of generation to get one.

    TechCrunch, Google buys 400 MW from Fervo (2 Sep 2026)

  7. the whole wait

    Every one of these waits outlasts several models

    Interconnection queues run four to seven years while model generations run months.

    a large power transformer

    a new hall, from site control to first power

    a heavy-duty gas turbine, ordered in 2026

    a connection to the transmission grid

    each blue tick: one more frontier model shipped while you wait

    LBNL, Queued Up: interconnection queues · Utility Dive, GE Vernova gas turbine backlog climbs to 116 GW (2026)

Where the wall is highest

The wall is not the same height everywhere. Whether a place is short of power depends on how much more its data centres want, how fast it can build, and above all how long it takes to connect what it builds. Some places have power to spare and are short of something else. Others have the plants on paper and can’t plug them in.

the world

415 → 945 TWh a year of data-centre electricity, 2024 → 2030

The IEA expects data centres to use more than twice as much electricity by 2030, slightly more than Japan uses today. The US and China account for nearly 80% of the growth. About half of it is expected to be met by renewables, and gas adds 175 TWh, mostly in the US.

IEA, Energy and AI, executive summary (2025)

data centres now expected by 2030, on one scale

  1. United States

    the wall is here

    ≈185 → ≈425 TWh, 2024 → 2030 · 2024 level worked out from the IEA’s increase and growth rate

    Data centres were 4.4% of US electricity in 2023; by 2028 they could be 6.7 to 12%.

    Building: A record 86 GW of new plants is planned for 2026, mostly solar and batteries.

    The catch: 1,312 GW of plants are waiting in queues to connect, and the median project built in 2025 waited more than five years. The power exists on paper; the connection is the wall.

    IEA, Energy and AI: energy demand from AI (2025) · Berkeley Lab, 2024 United States Data Center Energy Usage Report · Enerdata, US EIA forecasts record 86 GW of capacity additions in 2026 · Berkeley Lab, Queued Up: 2026 Edition

  2. China

    power to spare

    ≈103 → ≈278 TWh, 2024 → 2030 · 2024 level worked out from the IEA’s increase and growth rate

    China used 10,400 TWh in 2025. Its data centres’ whole increase to 2030 is under 2% of that.

    Building: More than 430 GW of wind and solar in 2025 alone, a record.

    The catch: Power is not China’s wall. Its wall is the chips, which export controls keep out of reach.

    IEA, Energy and AI: energy demand from AI (2025) · Gov.cn, China’s power consumption hits 10-trillion-kWh milestone in 2025 (17 Jan 2026) · Gov.cn, China’s newly installed wind and solar capacity up 22% in 2025 (12 Feb 2026)

  3. Europe

    tight

    ≈64 → ≈109 TWh, 2024 → 2030 · 2024 level worked out from the IEA’s increase and growth rate

    The smallest rise of the big regions: more than 45 TWh by 2030.

    Building: Renewables, slowly; grids were built for flat demand.

    The catch: Connection waits run to years in the main hubs, and Dublin and Amsterdam have both paused new data centres at points.

    IEA, Energy and AI: energy demand from AI (2025) · Norwegian SciTech News, Data centres’ demand for electricity (Nov 2025)

  4. Japan

    tight

    ≈19 → ≈34 TWh, 2024 → 2030 · 2024 level worked out from the IEA’s increase and growth rate

    From 19 TWh in 2024 to 57 to 66 TWh by 2034, by Wood Mackenzie’s estimate.

    Building: Mostly gas and coal around Tokyo and Osaka, where the data centres are.

    The catch: Reserve margins are above 15%, so no immediate shortage, but a new gas plant takes seven to ten years.

    IEA, Energy and AI: energy demand from AI (2025) · DCD, Data center energy consumption in Japan to triple by 2034 (Wood Mackenzie)

  5. India

    tight

    10–15 → 40–45 TWh, 2025 → 2030

    Capacity from about 1.2 GW in 2025 to almost 10 GW by 2030.

    Building: Fast-growing solar, with about $200bn of data-centre investment hoped for by 2030.

    The catch: The hubs are concentrated in a few states, whose grids take the whole load.

    Outlook Business, Data centre boom pushes India to rethink power planning (8 Apr 2026)

  6. Ireland

    the wall is here

    no comparable TWh forecast; the number that matters is below

    Data centres used 22% of Ireland’s metered electricity in 2024, up from 5% in 2015.

    Building: Since December 2025, a new data centre has to bring its own on-site generation and match at least 80% of its demand with new Irish renewables.

    The catch: The first country where the grid said no, and then wrote the price of yes.

    CSO, Data centres metered electricity consumption 2024 · CRU, decision on new electricity connection policy for data centres (Dec 2025)

  7. The Gulf

    power to spare

    no comparable TWh forecast; the number that matters is below

    Abu Dhabi’s planned 5 GW UAE–US AI campus; its first 200 MW goes live in 2026.

    Building: Nuclear from Barakah, solar and gas in the UAE; Saudi Arabia’s Humain targets 1.9 GW by 2030 on land with access to 14 GW of power.

    The catch: Cheap power and fast permits. The wall here is access to the chips, set by US export licences, and water for cooling in the desert.

    DCD, Companies behind UAE Stargate offer additional details · Blackridge Research, Middle East data centre projects 2026

  8. The Nordics

    power to spare

    no comparable TWh forecast; the number that matters is below

    Norway’s data centres could use 6 to 8 TWh by 2030.

    Building: Hydropower already makes over 90% of Norway’s electricity, with a surplus to spare.

    The catch: Among the least congested grids in Europe; the limit is distance from users, and a cold climate is a cooling bonus.

    Norwegian SciTech News, Data centres’ demand for electricity (Nov 2025)

Read down the stamps and a pattern shows. China and the Gulf have power to spare and are short of chips, which export rules decide. The US has the chips and the money, and its wall is the queue to connect. Ireland hit the wall first and turned it into a rule: bring your own power. For anyone choosing where to build, location is no longer a real-estate question. It decides which wall you meet.