OpenAI Stargate Abilene планирует строительство стоимостью около 32 млрд долларов на примерно 1 ГВт — цель на 2 года

Stay in the loop

Free daily newsletter & Telegram daily report

Join Telegram Channel

Executive Summary

Свежий транш данных привязывает масштаб и темп вычислений ИИ к реальным срокам и долларам. EpochAI очерчивает пять гиперскейлерских кампусов мощностью более 1 ГВт, ориентируясь на 2026 год, и оценивает расходы на ИИ-центры обработки данных в 2025 году около 300 млрд долларов (~1% ВВП США). Самостоятельно Stargate Abilene от OpenAI нацеливается на примерно 1 ГВт за примерно 2 года за ~32 млрд долларов, в то время как xAI утверждает, что Colossus 2 может достичь гигаватта примерно за 12 месяцев за счёт повторного использования оболочек и самозапуску локального питания на месте.

Вот реальная рабочая картина: сборки до 1 ГВт приходят за 1–4 года; размещение следует за электроэнергией, а не за близостью, потому что время генерации модели в 100 раз превосходит сетевой RTT (даже лунный ретранслятор не стал бы узким местом). Жидкое охлаждение обязательно, когда стойка размером 0,5 м² «пьёт» мощность 100 домов. Ранние этапы — привод от газовых турбин, затем добавление ветра/солнечной энергии после межсоединений, и лишь немногие страны могут разместить множество таких установок — 30 ГВт составляют около 5% генерации США, но около 90% в Великобритании.

Для команд, работающих над моделями, вывод таков: ожидается достаточная совокупная раскрутка мощности для поддержки примерно 5-кратного ежегодного роста на передовой обучении в течение двух лет, поэтому прямо сейчас планируйте многоступенчатые дорожные карты и региональную избыточность. Если вам нравятся амбициозные проекты вроде moonshots, проект Google «Suncatcher» рисует орбитальные TPU с линками 10 Тбит/с за примерно 810 долларов за кВт в год — но ближайшая цель — отслеживать реальные площадки через поля охлаждения и разрешения на виду.

Feature Spotlight

Feature: Gigawatt AI data centers race to 2026

Epoch AI’s analysis shows five 1+ GW AI data centers targeted for 2026, with cumulative AI DC spend >$300B (~1% US GDP). Power, not latency, dictates siting; xAI’s Colossus 2 aims GW scale in 12 months.

Multiple posts detail AI data centers scaling to 1+ GW with concrete timelines, costs, and power math; a dense infrastructure burst today across sources.

Jump to Feature: Gigawatt AI data centers race to 2026 topics

Table of Contents

🏗️ Feature: Gigawatt AI data centers race to 2026

Gigawatt data centers are hitting 1 GW in 1–2 years; observed range 1–4 years

xAI’s Colossus 2 targets 1 GW in ~12 months using shell reuse and early on‑site power

Only a few countries can host many >1 GW sites; 30 GW is ~5% US, ~2.5% China, ~90% UK

Power beats latency: model generation time is >100× network RTT

AI is ~1% of U.S. power today vs 8% for lighting and 12% for AC

Builds start with firm on‑site gas, then add grid renewables at scale

Many sites aim for 1–2 year paths to 1 GW despite heavy permitting

New Frontier Data Centers dataset and map launched for satellite‑based tracking

These campuses aren’t secret: satellite cooling arrays and permits expose progress

Design note: dense racks and liquid cooling define GW campus architecture


🧰 Agentic coding stacks and dev ergonomics

Berkeley’s DocETL debuts natural‑language→pipeline generator with hosted playground

GPT‑5‑Codex‑Mini throughput doubles; better for code‑heavy generations

RepoPrompt 1.5.28 adds Codex CLI provider and renders Codex reasoning traces

AG‑UI gets a Kotlin SDK to run agent UIs natively on Android/iOS/JVM

Agent evals highlight efficiency: GLM‑4.6 solved tasks with fewer steps

Conductor ships Plan Mode and a new copy‑path shortcut

Graphite’s gt modify now auto‑updates stacked PRs end‑to‑end

Cursor merges MCP server PR that unlocks ~8,000 web data sources

McPorter patch adds import of OpenCode configs for smoother CLI handoff

Yutori’s Scouts now learn from email replies to their reports


🧩 MCP interoperability and auth realities

Agent auth isn’t OAuth: call for short‑lived app keys and anomaly flags

Google publishes ‘Agent Tools & Interoperability with MCP’ white paper

Cursor merges MCP server PR unlocking 8,000+ web data sources

Groq, Docker and E2B host MCP hackathon with 200+ preinstalled tools

OSS “MCP alternative” in the works, promising backward compatibility

Community debate: “Does MCP suck?” recap surfaces DX and reliability gaps


⚙️ Serving speed and provider routing

Kimi K2 Thinking routing: SGLang on Atlas Cloud; vendors cite 0.3 s TTFT and ~140 TPS

Replit apps can now route to 300+ models via OpenRouter out of the box

GPT‑5‑Codex Mini doubles tokens/sec for code generation


📊 Leaderboards, eval variance, and efficiency scoring

Kimi flags 20+ pp benchmark drops via third‑party providers; urges official endpoint

Kimi K2 Thinking hits #2 open-source and ties #7 overall on Arena Text

Efficiency matters: GLM‑4.6 scores higher by finishing in fewer action steps

LisanBench: Kimi K2 Thinking lands 1928.6 Glicko‑2, between GPT‑5 and GPT‑5‑Mini

Artificial Analysis ranks LTX‑2 Pro #3 (image→video) and #7 (text→video)

Task length keeps rising: models double workable duration every ~7 months

Field report: K2 Thinking looping on Extended Connections despite official playground


🗣️ Speech and voice infra at scale

Meta open-sources Omnilingual ASR covering 1,600+ languages, up to 7B params

ElevenLabs Summit SF teases product announcements with enterprise partners


🎨 Generative media: Ketchup leaks, LTX‑2 results, editing tools

Google’s Nano‑Banana 2 shows up as “KETCHUP” in code, signaling a rename

Leaked “Ketchup” samples show strong instruction fidelity on visual puzzles

LTX‑2 Pro ranks #3 (image→video) and #7 (text→video); priced at $3.60/min 1080p

Replicate ships Reve Image Edit Fast at ~$0.01/output for spatially aware edits

Freepik Spaces adds camera‑angle controls on a collaborative canvas

Higgsfield Lipsync Studio demos emotion‑controlled lipsync across images and video

Creators highlight Firefly 5’s photo‑real output with film‑style prompts

Qwen‑Edit‑2509 Upscale LoRA targets photo restoration and detail recovery


🗂️ Data pipelines, retrieval, and NL→pipeline tools

Berkeley’s DocETL turns plain English into runnable data pipelines

UltraRAG 2.1 ships VisRAG, automated corpus server, and multimodal evals

Google Opal demo: one‑prompt agent chains research, tools, and Gemini 2.5 Pro

Parallel’s llms.txt turns any website into LLM‑friendly text via Extract API

Hornet positions a retrieval engine purpose‑built for agent workloads

Make.com AI connectors: Google Sheets → OpenAI tweet generator in minutes

TOON proposes a compact JSON replacement to cut prompt token spend


💼 Enterprise momentum and access programs

Gamma raises Series B at ~$2.1B; says it crossed ~$100M ARR

Kazakhstan brings ChatGPT Edu to 165 universities covering ~2.5M students

OpenAI offers one year of ChatGPT Plus free to eligible U.S. veterans

Replit apps can now use 300+ OpenRouter models out‑of‑the‑box

Kwai’s Kat Coder Pro joins OpenRouter, free for now; 73.4% SWE‑Bench Verified


🧠 Accelerators and parallelism notes

NVIDIA details Wide Expert Parallelism for MoE on GB200 NVL72, up to 1.8× per‑GPU

Nvidia reportedly pushes TSMC for +50% 3nm output to feed multi‑GW clusters

Blackwell NVFP4 kernel challenge launches; first task is NVFP4 GEMV

CoreWeave CEO says A100s are sold out across the market


🤖 Humanoids, autonomy, and logistics

XPENG says IRON humanoid targets mass production by 2026

DoorDash, Uber, Lyft flag bigger 2026 autonomy spend to scale robots and AVs

Waymo details driverless stack and safety; billions of miles in sim

NVIDIA Isaac Lab 2.3 hits GA to speed robot learning loops

Paper: Robot learns via a physical world model for fast action selection

‘Steve’ mod brings LLM agents into Minecraft for NL tasking


📚 Reasoning, spatial vision, and hallucination mitigation

Meta open-sources Omnilingual ASR for 1,600+ languages (up to 5,400 with one‑shot)

Real-time reasoning agents meet deadlines with AgileThinker’s dual-loop planning

VisAlign reduces VLM hallucinations by fusing a compact visual summary into text tokens

Visual Spatial Tuning (VST) claims SOTA spatial reasoning with 4.1M-sample corpus + RL

‘Jr. AI Scientist’ runs paper→hypothesis→experiments→draft, scores ≈5.75 vs. 3–4 baselines

NVIDIA’s Nemotron Nano V2 VL targets long docs/video with hybrid Mamba‑Transformer

ByteDance’s MIRA shows visual steps boost multimodal reasoning on conflict tasks


🛡️ Risk, labor impact, and market caution

Goldman flags dot‑com echoes in AI trade; capex ~$349B, spreads widen

Yale study: AI exposure shares stable since 2023; no unemployment spike

Debate over AI server “useful life” changes: audits back extensions, Amazon shortens

Developer trust in AI coding slips to ~60% from 70% in two years

Eric Schmidt: “Giving AI agency is a mistake” in near term

Field experiments: GenAI lifts online retail sales up to 16.3% via conversion gains

Germany’s software job postings fall below 60 (Feb ’20=100) after 2022 peak

On this page

Executive Summary
Feature Spotlight: Feature: Gigawatt AI data centers race to 2026
🏗️ Feature: Gigawatt AI data centers race to 2026
Gigawatt data centers are hitting 1 GW in 1–2 years; observed range 1–4 years
xAI’s Colossus 2 targets 1 GW in ~12 months using shell reuse and early on‑site power
Only a few countries can host many >1 GW sites; 30 GW is ~5% US, ~2.5% China, ~90% UK
Power beats latency: model generation time is >100× network RTT
AI is ~1% of U.S. power today vs 8% for lighting and 12% for AC
Builds start with firm on‑site gas, then add grid renewables at scale
Many sites aim for 1–2 year paths to 1 GW despite heavy permitting
New Frontier Data Centers dataset and map launched for satellite‑based tracking
These campuses aren’t secret: satellite cooling arrays and permits expose progress
Design note: dense racks and liquid cooling define GW campus architecture
🧰 Agentic coding stacks and dev ergonomics
Berkeley’s DocETL debuts natural‑language→pipeline generator with hosted playground
GPT‑5‑Codex‑Mini throughput doubles; better for code‑heavy generations
RepoPrompt 1.5.28 adds Codex CLI provider and renders Codex reasoning traces
AG‑UI gets a Kotlin SDK to run agent UIs natively on Android/iOS/JVM
Agent evals highlight efficiency: GLM‑4.6 solved tasks with fewer steps
Conductor ships Plan Mode and a new copy‑path shortcut
Graphite’s gt modify now auto‑updates stacked PRs end‑to‑end
Cursor merges MCP server PR that unlocks ~8,000 web data sources
McPorter patch adds import of OpenCode configs for smoother CLI handoff
Yutori’s Scouts now learn from email replies to their reports
🧩 MCP interoperability and auth realities
Agent auth isn’t OAuth: call for short‑lived app keys and anomaly flags
Google publishes ‘Agent Tools & Interoperability with MCP’ white paper
Cursor merges MCP server PR unlocking 8,000+ web data sources
Groq, Docker and E2B host MCP hackathon with 200+ preinstalled tools
OSS “MCP alternative” in the works, promising backward compatibility
Community debate: “Does MCP suck?” recap surfaces DX and reliability gaps
⚙️ Serving speed and provider routing
Kimi K2 Thinking routing: SGLang on Atlas Cloud; vendors cite 0.3 s TTFT and ~140 TPS
Replit apps can now route to 300+ models via OpenRouter out of the box
GPT‑5‑Codex Mini doubles tokens/sec for code generation
📊 Leaderboards, eval variance, and efficiency scoring
Kimi flags 20+ pp benchmark drops via third‑party providers; urges official endpoint
Kimi K2 Thinking hits #2 open-source and ties #7 overall on Arena Text
Efficiency matters: GLM‑4.6 scores higher by finishing in fewer action steps
LisanBench: Kimi K2 Thinking lands 1928.6 Glicko‑2, between GPT‑5 and GPT‑5‑Mini
Artificial Analysis ranks LTX‑2 Pro #3 (image→video) and #7 (text→video)
Task length keeps rising: models double workable duration every ~7 months
Field report: K2 Thinking looping on Extended Connections despite official playground
🗣️ Speech and voice infra at scale
Meta open-sources Omnilingual ASR covering 1,600+ languages, up to 7B params
ElevenLabs Summit SF teases product announcements with enterprise partners
🎨 Generative media: Ketchup leaks, LTX‑2 results, editing tools
Google’s Nano‑Banana 2 shows up as “KETCHUP” in code, signaling a rename
Leaked “Ketchup” samples show strong instruction fidelity on visual puzzles
LTX‑2 Pro ranks #3 (image→video) and #7 (text→video); priced at $3.60/min 1080p
Replicate ships Reve Image Edit Fast at ~$0.01/output for spatially aware edits
Freepik Spaces adds camera‑angle controls on a collaborative canvas
Higgsfield Lipsync Studio demos emotion‑controlled lipsync across images and video
Creators highlight Firefly 5’s photo‑real output with film‑style prompts
Qwen‑Edit‑2509 Upscale LoRA targets photo restoration and detail recovery
🗂️ Data pipelines, retrieval, and NL→pipeline tools
Berkeley’s DocETL turns plain English into runnable data pipelines
UltraRAG 2.1 ships VisRAG, automated corpus server, and multimodal evals
Google Opal demo: one‑prompt agent chains research, tools, and Gemini 2.5 Pro
Parallel’s llms.txt turns any website into LLM‑friendly text via Extract API
Hornet positions a retrieval engine purpose‑built for agent workloads
Make.com AI connectors: Google Sheets → OpenAI tweet generator in minutes
TOON proposes a compact JSON replacement to cut prompt token spend
💼 Enterprise momentum and access programs
Gamma raises Series B at ~$2.1B; says it crossed ~$100M ARR
Kazakhstan brings ChatGPT Edu to 165 universities covering ~2.5M students
OpenAI offers one year of ChatGPT Plus free to eligible U.S. veterans
Replit apps can now use 300+ OpenRouter models out‑of‑the‑box
Kwai’s Kat Coder Pro joins OpenRouter, free for now; 73.4% SWE‑Bench Verified
🧠 Accelerators and parallelism notes
NVIDIA details Wide Expert Parallelism for MoE on GB200 NVL72, up to 1.8× per‑GPU
Nvidia reportedly pushes TSMC for +50% 3nm output to feed multi‑GW clusters
Blackwell NVFP4 kernel challenge launches; first task is NVFP4 GEMV
CoreWeave CEO says A100s are sold out across the market
🤖 Humanoids, autonomy, and logistics
XPENG says IRON humanoid targets mass production by 2026
DoorDash, Uber, Lyft flag bigger 2026 autonomy spend to scale robots and AVs
Waymo details driverless stack and safety; billions of miles in sim
NVIDIA Isaac Lab 2.3 hits GA to speed robot learning loops
Paper: Robot learns via a physical world model for fast action selection
‘Steve’ mod brings LLM agents into Minecraft for NL tasking
📚 Reasoning, spatial vision, and hallucination mitigation
Meta open-sources Omnilingual ASR for 1,600+ languages (up to 5,400 with one‑shot)
Real-time reasoning agents meet deadlines with AgileThinker’s dual-loop planning
VisAlign reduces VLM hallucinations by fusing a compact visual summary into text tokens
Visual Spatial Tuning (VST) claims SOTA spatial reasoning with 4.1M-sample corpus + RL
‘Jr. AI Scientist’ runs paper→hypothesis→experiments→draft, scores ≈5.75 vs. 3–4 baselines
NVIDIA’s Nemotron Nano V2 VL targets long docs/video with hybrid Mamba‑Transformer
ByteDance’s MIRA shows visual steps boost multimodal reasoning on conflict tasks
DeepEyesV2 pitches an agentic multimodal stack; short explainer and paper link
🛡️ Risk, labor impact, and market caution
Goldman flags dot‑com echoes in AI trade; capex ~$349B, spreads widen
Yale study: AI exposure shares stable since 2023; no unemployment spike
Debate over AI server “useful life” changes: audits back extensions, Amazon shortens
Developer trust in AI coding slips to ~60% from 70% in two years
Eric Schmidt: “Giving AI agency is a mistake” in near term
Field experiments: GenAI lifts online retail sales up to 16.3% via conversion gains
Germany’s software job postings fall below 60 (Feb ’20=100) after 2022 peak