★
Menu
Home How It's Built Watchlist The Stack
Hardware ChipsCompare Chips Servers & SystemsNetworking MemoryFoundries & Packaging Cooling
Facilities Data CentersPower ProjectsCloud & GPU Compute
Intelligence AI ModelsSoftware AI ProductsBenchmarks Research PapersDatasets Concepts
Ecosystem CompaniesStocks PeopleTimeline TrendsFuture of AI
Learn News Ask the Data AI Agents
About About us Methodology Data & API Disclaimer
Options
DATA & API

Free read-only JSON access to the site's cached data.

Dark mode

Guided View
New to chips, data centers and model jargon? We explain key terms as you browse, in plain English. Same data, with the help built in.

Expert View
You already know the stack. Just the data: clean, fast and compact, with no extra explanations. This is the default view.

Interface language
‹ Back
Challenger siliconThe next platform generationNew fabrics, memory and lightNew powerBeyond the transformer

Challenger silicon

8

Shipping or credible alternatives to the incumbent GPUs — each betting on a different corner of the workload.

CW Cerebras WSE-3Wafer-scale compute with tokens-per-second records — and an IPO on file. Shipping Chip GL Groq LPUDeterministic inference at famous speed; Saudi-scale deployment deals. Shipping Chip ES Etched SohuThe all-in transformer ASIC — if it ships, the specialization era begins. In development Chip DM d-Matrix CorsairIn-memory compute aimed at cheap interactive serving. Ramping Chip FR FuriosaAI RNGD150 W LLM inference; turned down Meta to stay independent. Ramping Chip RR Rebellions REBELKorea's merged champion, fabbed by Samsung. In development Chip TB Tenstorrent BlackholeJim Keller's open RISC-V stack, sold and licensed. Shipping Chip QC Qualcomm Cloud AI / AI200Mobile-silicon economics re-entering the rack in 2026. Announced Chip

The next platform generation

5

Already announced, already contracted — the hardware the 2026–27 buildout will install.

NR NVIDIA RubinHBM4, the Vera CPU, NVL144 racks — the incumbent's next lap. Ramping Chip NR NVIDIA Rubin CPXA cheaper GDDR7 chip just for long-context prefill — inference splits in two. Announced Chip AI AMD Instinct MI400 / MI450432 GB of HBM4 and the OpenAI 6 GW deal riding on it. Ramping Chip AT AWS Trainium33 nm and a claimed 4x — the custom-silicon bet compounds. Announced Chip GT Google TPU v7 (Ironwood)9,216-chip pods for the inference age, with Anthropic as anchor. Ramping Chip

New fabrics, memory and light

7

The interconnect and memory transitions that decide whether open ecosystems can match proprietary racks.

UA UALinkThe open answer to NVLink — 1,024 accelerators per pod, everyone-but-NVIDIA. Technology UE Ultra EthernetEthernet rebuilt for AI; spec 1.0 shipped, silicon arriving. Technology CP Co-Packaged OpticsLight moves onto the switch package; ~3x interconnect efficiency. Technology LI LightmatterPhotonic interposers — light inside the package itself. Company CA Celestial AIOptical links to disaggregated memory — an exit from the HBM wall. Company HB HBM42,048-bit interfaces and foundry logic dies; memory and logic start merging. Technology CX CXLPooled memory over PCIe wiring — slow burn, big if it lands. Technology

New power

7

The generation wave AI money is financing — targets, not certainties, and labeled as such.

SM Small Modular Reactors (SMRs)Factory-built reactors sized for campuses; every date is a target. Technology OK OkloFast reactors with a 12 GW Switch framework and a build-own-operate model. Company XE X-energyAmazon-backed pebble-bed reactors targeting ~2030. Company KH Kairos Hermes / Google FleetGoogle's 500 MW fleet order — first Gen IV construction permits in the US. Under construction (demo) Power Project TN TerraPower Natrium (Kemmerer)The furthest-along advanced reactor build in America. Under construction Power Project FC Fuel CellsBloom-style bridge power while the grid queue clears. Technology HC Homer City Energy Campus4.5 GW of gas-for-AI on a demolished coal site — the biggest such plan. Announced Power Project

Beyond the transformer

5

Architecture bets that would redraw the hardware map if they win.

SS State-Space Models (Mamba)Linear-time sequence models — production hybrids already ship. Concept WM World ModelsLearning reality's dynamics, not just text — DeepMind, NVIDIA and World Labs all in. Concept M( Marble (World Labs)Fei-Fei Li's spatial-intelligence bet, now a product. Current AI Model G3 Genie 3Playable generated worlds — training grounds for agents and robots. Current AI Model TT Test-time Compute & ReasoningThe scaling axis that already rewrote inference demand once. Concept

House rule: every date on the linked records is the maker's stated target, labeled as such. This page tracks claims — it does not endorse them, and it is never investment advice.