AI Chips · Comparison

Compare the AI Chips, Side by Side

Every brand we cover in one table, using each company’s own published figures. Chips are not directly comparable unless the numbers are stated the same way, so read the cautions below before drawing conclusions.

Stuck at any point? Ask an AI about this page →
11 brandsCompany figures onlyChecked September 2026
How to use this table. Each row repeats what that company publishes, and links to the full guide where every figure carries its source. Where a company publishes no price, the row says so rather than guessing.

Current flagship chips

BrandCurrent flagshipMemory per chipAI compute per chipMade byPublished price
NVIDIARubin (Vera Rubin NVL72)288 GB HBM450 PFLOPS NVFP4 inference; 35 trainingTSMCNot published; rentable from about $5 per H100 GPU-hour (AWS Capacity Blocks)
AMDInstinct MI455X (Helios rack)432 GB HBM440 PFLOPS FP4; 20 FP8 (dense/sparse not stated)TSMCNot published; rentable from under $2 per MI300X GPU-hour (Vultr, pre-emptible)
IntelGaudi 3128 GB HBM2e1,678 TFLOPS BF16/FP8 (white paper); 1,835 in Intel's Hot Chips slidesTSMC (5 nm)$125,000 list for an 8-chip kit (Intel, 2024); IndiaAI ₹153/hour per card before GST
Google TPUIronwood (TPU7x); TPU 8t/8i coming192 GiB HBM4,614 TFLOPS FP8 (Google, "peak")Google with Broadcom; foundry not disclosed$12.00 per chip-hour on demand (Google Cloud list price)
Amazon TrainiumTrainium3 (Trn3 UltraServer)144 GB HBM3e2.52 PFLOPS FP8 (AWS)Annapurna Labs designs; foundry not disclosedTrn3 not published; Trainium2 $2.235 per chip-hour (AWS Capacity Block)
Microsoft MaiaMaia 200216 GB HBM3eOver 10 PFLOPS FP4; over 5 PFLOPS FP8 (Microsoft)TSMC (3 nm)Not for sale or rent (used for Microsoft services)
Meta MTIAMTIA 300 (MTIA 400 next)216 GB HBM3E1.2 PFLOPS FP8 (as reported from Meta's roadmap)Meta with Broadcom; TSMC for earlier chipsNot for sale or rent (Meta internal use only)
AppleM5 Ultra (Mac Studio)Up to 512 GB unified memory, 1.2 TB/sNot published as TOPS (last figure: M4, 38 TOPS)Apple designs; TSMC makesMac Studio with M5 Ultra from $5,499
CerebrasWSE-3 (CS-3); CS-4 shipping44 GB SRAM on the wafer, 21 PB/s125 PFLOPS per WSE-3; 250 per WSE-3 TurboCerebras designs; TSMC 5 nmCloud: gpt-oss-120b $0.35 in / $0.75 out per million tokens
GroqLPU (gen 1); NVIDIA Groq 3 LPX under licence230 MB SRAM per chip, 80 TB/s750 TOPS INT8 / 188 TFLOPS FP16 per chipGroq design; 14 nm (GlobalFoundries per press)Cloud: GPT OSS 120B $0.15 in / $0.60 out per million tokens
Huawei AscendAscend 910C; 950 series from 2026950DT: 144 GB own HBM, 4 TB/s950: 1 PFLOPS FP8 / 2 PFLOPS MXFP4HiSilicon; factory not disclosed (press: SMIC)Not published by the company

Who makes what, and what each is good at

BrandWho manufactures itReal strength
NVIDIATSMC (chips) with partners for packaging and assemblyThe whole stack: chips, networking, racks and CUDA software
AMDTSMC (chiplets), with partners for packaging and assembly432 GB of memory per chip, open software, and gigawatt-scale AI-lab deals
IntelIntel's own fabs for its processors; TSMC for Gaudi 3Xeon runs inside NVIDIA's DGX systems, and Intel owns chip factories in the US and Europe
Google TPUGoogle designs; Broadcom co-develops and supplies; foundry not disclosed9,216-chip pods, published per-chip prices, and a million-chip Anthropic deal
Amazon TrainiumAnnapurna Labs (Amazon) designs; foundry not disclosedAnthropic and OpenAI deals, and a chips business past a $25 billion run rate
Microsoft MaiaMicrosoft designs; TSMC manufactures216 GB of memory per chip, and it already runs GPT-5.2 and Copilot
Meta MTIAMeta designs with Broadcom; TSMC made earlier generationsHundreds of thousands of chips in service and a new chip every six months
AppleApple designs; TSMC makes, including in ArizonaAI hardware in every iPhone, iPad and Mac, with one shared pool of memory
CerebrasCerebras designs; TSMC makes the wafers (5 nm)One chip the size of a dinner plate, with 44 GB of memory on the chip itself
GroqGroq designed; 14 nm first chip; Samsung makes Groq 3 for NVIDIAAll memory on the chip, scheduled in advance by the compiler, for very fast answers
Huawei AscendHiSilicon designs; Huawei does not name its factories (press: SMIC; stockpiled TSMC dies)Scale over single-chip speed: thousands of chips joined into one system

Five cautions before you compare

  1. Dense against dense. A figure marked “with sparsity” is usually double the plain one.
  2. Same precision. FP4 numbers are roughly double FP8 numbers on the same chip, and FP4 itself comes in two incompatible forms.
  3. Same unit. Some companies quote one chip, some a server of eight, some a rack of 72.
  4. Memory matters as much as speed. Capacity decides whether a model fits; bandwidth decides how fast it answers.
  5. Prices are rarely comparable. Some companies sell chips, some rent them by the hour, and several publish nothing at all.

Each of these is explained, with the vendors’ own wording, in How to read a chip spec sheet. For the words themselves, see the glossary.

Ask an AI about this page

Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.

ChatGPTClaudeGeminiPerplexityGrok

Nothing is sent from here. The link carries only this page’s title and address.