Tum Claude ya ChatGPT kholte ho, sawaal type karte ho, do second mein jawab mil jaata hai. Din mein pachaas baar karte ho aur assume karte ho ki ye aur tez aur sasta hota jaayega — jaise software hamesha se hota aaya hai. Reasonable expectation. Ek problem hai.

Har ek query physical hardware pe chalti hai jo bijli khaata hai jaise ek chhota sheher khaata hai. Pichle do hafton mein humne alag-alag pieces cover kiye — Amazon ka $200 billion capex bet aur usse paida hua cash flow squeeze, NAACP ka xAI pe case Mississippi mein 27 bina permit ke gas turbines ke liye, TSMC ka record Q1 jisne confirm kiya ki uska 61% revenue ab un chips se aata hai jinhe kahin na kahin plug in karna padta hai. Har story apne aap mein samajh aati thi. Saath mein, ye ek alag kahani sunati hain — ki 2028 mein kaun si companies actually tumhe serve karengi aur kaun si nahi.

Thesis: Power naya moat hai

AI race ki conventional framing intelligence ke baare mein hai — kiske paas best model hai, highest benchmark score hai, sabse capable agent hai. Lekin model quality converge ho rahi hai. Electricity ka access nahi.

Tom's Hardware ki ek analysis ke mutabiq jo 3 April, 2026 ko publish hui, utilities aur grid operators ne is saal ke liye planned ~140 bade US data center projects mein se lagbhag aadhe delay ya cancel kar diye. Paison ki ya demand ki kami ki wajah se nahi — electrical grid ki wajah se. Transformer lead times paanch saal tak pahunch gayi hain. Naye nuclear plants mein ek dashak lagta hai. Microsoft ke CEO Satya Nadella ne January 2026 mein maana ki company ke paas GPUs inventory mein pade hain kyunki unhe plug in karne ki jagah nahi hai.

Tum ek better model chhe mahine mein ship kar sakte ho. Power plant nahi kar sakte.

Vendor Power Scorecard

Yahan se baat practical hoti hai. Agar tum apni team ke liye AI tools choose kar rahe ho — coding assistant, API provider, agents ke liye platform — to April 2026 tak ka power landscape kuch aisa dikhta hai, organized by ki kisne actually generation capacity secure ki vs. kaun abhi bhi scramble kar raha hai.

Tier 1 — Power-secured (generation capacity own ya lease karti hain long-term contract ke under)

  • Microsoft ne September 2024 mein Three Mile Island ka Unit 1 reactor (835 MW) restart karne ke liye 20 saal ka deal sign kiya. Company ke paas multiple regions mein long-term renewable PPAs bhi hain.
  • Amazon ne March 2024 mein announce kiye gaye deal mein Susquehanna plant se 2042 tak ke liye 1.9 GW nuclear power lock in kiya. Amazon seedha solar aur wind farms bhi own karta hai.
  • Google ne October 2024 mein Kairos Power ke saath 500 MW small modular reactors ke contracts sign kiye, Kairos ka target late 2020s mein delivery hai.

In companies ne current crunch se pehle saalon aur billions lagakar electrons secure kiye. Ye head start ab ek structural advantage hai jo koi model improvement cross nahi kar sakti.

Tier 2 — Power-dependent (grid se kharid rahe hain, permits ke liye compete kar rahe hain)

  • Meta aggressively data centers banata hai lekin primarily grid power aur renewable energy credits pe depend karta hai. Tier 1 ke scale ka koi announced generation ownership nahi.
  • Oracle tezi se expand kar raha hai bina disclosed long-term generation agreements ke jo upar wale nuclear deals ke barabar hon.

Dono bana sakte hain. Koi bhi fuel control nahi karta.

Tier 3 — Power-desperate (corners cut kar rahe hain)

  • xAI ne Southaven, Mississippi mein apni Colossus 2 site pe bina air permits ke 27 gas turbines install kar diye. NAACP ne 14 April ko Clean Air Act ke under case kiya, estimate ke mutabiq saalana 1,700+ ton nitrogen oxides ek aise region mein pump ho rahe hain jo pehle se federal smog standards fail karta hai. Jab power ki desperation ek aisi community se milti hai jo aur pollution afford nahi kar sakti — ye jugaad nahi, ye laaparwaahi hai.

Futurum Group ki ek analysis jo 11 April, 2026 ko publish hui, usne paaya ki top five hyperscalers ab apne operating cash flows ka lagbhag 100% capex pe kharch kar rahe hain — jo 10 saal ke average 40% se upar hai. Ye concentration tiers ke beech gap ko aur tez karti hai. Jinke paas secured generation hai wo build karte reh sakte hain. Jinke paas nahi hai unke saamne ek ceiling hai jo sirf capital se nahi uthti.

Tumhare stack ke liye iska matlab

Jab tum 2027-2028 ke liye AI vendors evaluate karo, apni checklist mein ye sawaal add karo:

Kya provider generation capacity own ya lease karta hai? Sirf "renewable energy credits" nahi — actual megawatts long-term contract ke under. 10+ saal ke PPAs matter karte hain. Sirf RECs grid crunch ke waqt lights on hone ki guarantee nahi dete.

Unke data centers kahan hain? Jin regions mein grid headroom hai (US Southeast ke kuch hisse, Nordics, Canada ke kuch hisse) wo zyada reliably requests serve karenge power-constrained corridors (Northern Virginia, Texas ke kuch hisse, Netherlands) ke comparison mein. Apne vendor se pucho ki tumhara traffic kis region se route hota hai.

Unka capex-to-cash-flow ratio kya hai? Jo company apne operating cash flow ka 95% infrastructure pe kharch kar rahi hai wo ek bure quarter se capacity expansion cut karne ke kareeb hai. Multi-year contract sign karne se pehle last do earnings calls check karo.

Energy pe regulatory compliance ka track record hai? Aaj ke bina permit ke turbines kal overall operational risk signal karte hain. Agar ek vendor air permits pe corners cut karta hai, to pucho aur kya skip kar raha hai.

Duniya ka sabse smart model bekar hai agar usse chalane wali company lights on nahi rakh sakti. Sabse simple model jiske peeche nuclear plant hai wo 2030 mein bhi tumhare queries ka jawab dega.

AI race intelligence ke baare mein honi chahiye thi. Bottleneck bijli ka bill hai.