Every AI Chip · Tracked
Every AI chip from every manufacturer and every foundry. Specs, power, cloud pricing, known buyers, and live supply tightness · cross-linked to foundries, memory suppliers, racks, datacenters, and the models trained on them.
- Weighted tightness68
- Lead time multiplier2.09×
- Capex accelerator1.85×
- Frontier holdout premium+60
Tier Ladder
Frontier · Mainstream · Specialized · Legacy · assessed quarterly
Current-generation flagship silicon shipping at scale to hyperscalers. The chips training the biggest models right now.
Current-gen non-flagship or last-gen flagship still deployed widely. The workhorses of the AI economy.
Wafer-scale, custom hyperscaler silicon, or experimental architectures serving niche workloads.
Older chips being phased out or already past EOL. Still used in research clusters and price-sensitive workloads.
Performance per Watt
FP16 TFLOPs vs TDP · log scale · bubble size = HBM capacity
Up-and-to-the-left wins. Chips above the 4 T/W line are the efficiency frontier — that's where the power-budget game is won at hyperscale. Wafer-scale and multi-die superchips sit top-right because they trade watts for raw throughput.
Chip Demand Ladder
How tight is supply right now · hand-curated from order books + earnings calls
GB200GB300B200
Ascend 910C
WSE-3B300
MI355X
Trainium2MI325X
Maia 100H200
TPU v7
LPUTPU v6e
MTIA v2MI300XH100
Gaudi 3TPU v5pA100Supply notes
- GB200Sold out through 2027 · Oracle, Microsoft, Meta competing
- GB300xAI Memphis + Stargate phase 2
- B200CoWoS-L packaging capacity is the bottleneck
- Ascend 910CChina export-ban alternative · SMIC N7 constrained
- WSE-3G42 Condor Galaxy absorbing most wafers
- B300Ultra variant ramping Q4 2025
This is the per-chip preview of the AI Compute Demand Index landing on /compute · unbounded above 100, goes to 250+ when the market is on fire. Updated weekly from hyperscaler order books and earnings disclosures.
Best for
Six curated shortlists computed from the live spec data
Most raw FP16 throughput
Top 5 by FP16/BF16 TFLOPs · ignore price and power · pure compute
Most efficient (TFLOPs/W)
Best FP16 TFLOPs per Watt · the chips that make datacenter power budgets work
Most HBM capacity
Biggest memory for fitting 400B+ parameter models in a single node
Cheapest cloud rental
Lowest $/h in public clouds · best for quick experiments
Current flagship chips
The chips training the biggest frontier models right now
Full leaderboard
Sorted by Gecko score · composite of tier, recency, adoption, perf density
| # | Chip | Maker | Score |
|---|---|---|---|
| 1 | NVIDIA | 94 | |
| 2 | NVIDIA | 86 | |
| 3 | 83 | ||
| 4 | AWS | 78 | |
| 5 | NVIDIA | 75 | |
| 6 | AMD | 75 | |
| 7 | Huawei | 75 | |
| 8 | 71 | ||
| 9 | NVIDIA | 70 | |
| 10 | AMD | 68 | |
| 11 | AMD | 65 | |
| 12 | Cerebras | 60 | |
| 13 | Intel | 57 | |
| 14 | NVIDIA | 55 | |
| 15 | NVIDIA | 55 | |
| 16 | Meta | 52 | |
| 17 | 49 | ||
| 18 | Microsoft | 42 | |
| 19 | Groq | 34 | |
| 20 | NVIDIA | 31 |
By manufacturer
10 companies · 3 foundries · 2 countries
Frequently asked
Pulled from the live dataset · schema-ready for AEO
Which AI chip has the most FP16 TFLOPs right now?
NVIDIA GB300 (Grace Blackwell Ultra) leads dense FP16 throughput at ~5000 TFLOPs per superchip, followed by GB200 at 4500 TFLOPs. The Cerebras WSE-3 reaches 125000 TFLOPs per wafer but is a fundamentally different architecture · a single wafer contains what would otherwise be ~50 GPUs.
Which AI chip is most efficient per Watt?
Google TPU v7 (Ironwood) and NVIDIA B200 lead FP16 TFLOPs per Watt among shipping chips. Trillium (TPU v6e) is close behind. Inference-first designs (Ironwood, Groq LPU) tend to dominate perf/Watt because they skip training-specific circuitry.
Who manufactures AI chips besides NVIDIA?
10 companies ship production AI accelerators tracked on BenchGecko · NVIDIA, AMD, Google, AWS, Intel, Microsoft, Meta, Huawei, Cerebras, Groq, and more. NVIDIA dominates third-party sales but hyperscalers are shipping more custom silicon (Trainium2, TPU v7, Maia 100, MTIA) to reduce dependence.
Which fabs make AI chips?
90% of tracked chips are fabbed at TSMC. GlobalFoundries handles Groq's 14nm LPU. SMIC fabricates Huawei's Ascend 910C at their N7 node as the primary answer to the U.S. export ban. TSMC advanced nodes (N5/N4/N3) are the bottleneck for current-gen NVIDIA, AMD, AWS, Google, and Microsoft silicon.
How do you decide which chip is "frontier" tier?
Frontier means current-generation flagship, shipping in 2025-2026, widely ordered by hyperscalers. Mainstream means current-gen non-flagship or last-gen flagship still deployed (H100, H200, MI300X). Specialized means wafer-scale, LPU, MTIA and other custom designs. Legacy means EOL or phasing out (A100, V100). Tiers are reviewed quarterly.
Where does BenchGecko get chip data from?
Manufacturer datasheets for specs, SEC 10-K filings and earnings transcripts for buyer disclosures, cloud provider API pricing for $/hour, public press releases for release dates. Every chip detail page includes source links. No paywalled data · no vendor-sponsored rankings · everything is reproducible.
See also
Keep exploring the compute graph