The pricing layer of the tracker: 63 chips — 51 with credible public prices — 150 dated, source-linked observations, all shown in USD with the as-reported figure underneath. The question this page exists to answer: how far apart are Chinese chips and their global rivals, in dollars?
Dataset updated 2026-08-27 · every observation source-linked
131 TFLOPS/$1kcheapest compute on the tracker — AMD Instinct MI300X at ~$10K
29 TFLOPS/$1kbest Chinese-vendor value — Ascend 910C at ~$27.6K
~$9.7K vs ~$16.5Kmedian accelerator price, China-market (12) vs global (8)
11price observations on Nvidia H20 — launch to halt
The gap
China vs the global frontier
The spread of FP16 compute per dollar on each side, per generation — every qualifying chip, not just the leader. Gap compares each side’s best. Sides split by vendor nationality — China-market variants of US chips (H20, L20) sit on neither side.
FP16 compute per dollar: Chinese vendors sit about 1–6× behind the global frontier, depending on which chips you pair.
Why Nvidia still dominates. Nothing here says AMD builds the best AI chip — the ranking is peak-FP16-per-dollar and nothing else. AMD tops it by quoting large FP16 figures and pricing aggressively (our own MI325X note records “weak customer interest and room for price cuts”); Nvidia wins where it matters — CUDA and the software stack, real-world utilization of peak, NVLink networking, and whole-system delivery. There is also a price-observation bias: Nvidia publishes no list prices, so its figures skew toward scarcity street prices ($25–40K H100), while AMD's best observations skew toward negotiated volume deals ($10K MI300X via Microsoft/Citi). Deployments tell the story: Nvidia holds the overwhelming share of AI compute; AMD's share is real but small.
Why domestic chips still sell. The gap compares quoted prices in markets that export controls keep apart: the MI300X's ~$10K is a US-market contract price a Chinese buyer cannot legally pay. When US silicon does reach China, scarcity repricing closes most of the gap — the H800's post-ban China street price ran $35–70K per card, putting it in the same ~15–25 TFLOPS/$1k band as the best domestic parts. And buyers don't optimize peak FP16 per dollar: the H20 sold over a million units in 2024 at 11–15 TFLOPS/$1k — worse value than any chip above — because inference is memory-bound and supply security, after two H20 bans in 18 months, is worth a premium. State procurement (China Mobile's >RMB 1B Kunlunxin orders) does the rest.
Ratios compare each side's best value (most FP16 TFLOPS per $1,000) within a generation, priced at each chip's latest MARKET observation (street, contract, or list) — analyst models are used only where no market price exists, and marked “estimate” with a confidence dot. A ratio of 3× means the best global chip delivers three times more peak FP16 per dollar; real-world gaps are smaller where software, memory, and interconnect favor the global part — and larger where export controls inflate China-market pricing. Channel price spreads (e.g. Ascend 910C's ¥80k–200k quotes) can move a ratio substantially. Chips and gaps marked * are priced by estimate where no market observation exists — true of the entire 2025–2026 global side: neither MI355X nor B300 has a market per-card price yet, only analyst models.
The generations
Every chip, generation by generation
One dot per market-priced accelerator; full tables on the category pages.
Every market-priced AI-server accelerator, one dot per chip: columns are generations (newest first), height is the latest market price in USD at observation-year FX (log scale), red = Chinese vendor, gray = global. Hover any dot for its as-reported price; click through to the chip page. 15 tracked accelerators have no market-priced observation and are omitted. Full tables — with every observation, source, and unpriced chip — live on the category pages.
Ranked by peak FP16 per dollar — one axis, not a verdict: AMD tops paper FLOPs-per-dollar; Nvidia wins deployments on software, utilization, and systems. Leaderboard first, then every priced accelerator on one map.
Longer bar = more peak FP16 compute per dollar. This measures one narrow thing — quoted dense-FP16 throughput divided by the latest per-unit price (USD at observation-year average FX) — and deliberately ignores memory bandwidth, interconnect, software ecosystem, and real-world utilization, which is why AMD parts top raw FLOPs-per-dollar while Nvidia dominates actual deployments. Bars mix generations by default: an older chip at a discounted 2026 street price can outrank a new flagship at launch price — check each bar's generation tag, or filter to one generation for a like-for-like view. Datacenter class only; consumer gaming cards never rank here. Bar color: red = China-vendor, gray = global.
Metric:
Each dot is one chip: farther right = more compute, higher = more expensive (both scales stretch logarithmically so a cheap edge chip and a $30,000 flagship fit on one picture — equal steps multiply, they don't add). Bottom-right dots deliver the most compute per dollar; see the ranked bars above for the same story as a plain leaderboard. Point size = source confidence; red = China-market, gray = global. All prices in USD at the observation year's average exchange rate (the native figure shows on hover).
Compute per rental dollar 12 rates
1Ascend 910B1129 TFLOPS/$1hr$0.28/card-hr (derived) · per 8-card server-month (SMM Yangtze River Delta 91…
2Ascend 910B985 TFLOPS/$1hr$0.32/card-hr (derived) · per 8-card server-month (SMM West China 910B-64G av…
3Nvidia B200594 TFLOPS/$1hr$4/card-hr · per GPU-hour (cloud on-demand, 2025-07)
4AMD Instinct MI325X581 TFLOPS/$1hr$2/card-hr · per GPU-hour (cloud on-demand)
Longer bar = more FP16 compute per $1/hour of rental — the rental twin of the leaderboard above. Monthly 8-card machine leases are normalized to per-card-hour (÷8 cards ÷730 hours) and flagged “derived”; wholesale leases may bundle hosting and power, so treat them as a floor on rental value. CPUs are excluded (no FP16 figure). The rental market’s striking feature: wholesale 910B leasing delivers the most compute per rental dollar anywhere — while the same chip at official retail rates delivers the least.
Trends
Where prices are moving
Chips with three or more dated per-unit observations, ordered by total change — the memory supercycle, export-control discounts, and generational price cuts all show up here. Indexed to each chip's first observation; every figure lives on the chip's page with sources.
Each line is one chip's per-unit price observations over time, converted to USD at each observation year's average FX and indexed to its first observation (= 100). Hover a line or legend chip to highlight it; click legend chips to add or remove chips; the dashed line marks unchanged price. Analyst estimates and reseller listings are excluded — every point is a dated, sourced market observation (hover any dot). Chip pages carry the full history.
Every spec and price carries a source link and a date. Datacenter chips rarely have official list prices, so figures are what the market reported — launch MSRP, street, contract — and every observation is kept, because the trajectory is the story.
All prices display in USDat the observation year's average exchange rate (RMB per USD; HKD at 7.8); the original figure and rate print under each price. Confidence: green dot — vendor-official or multiple independent reports; amber — single major report; gray — estimate or lone secondary source. Raw research batches are archived under docs/chip-research/.
Prices from vendor listings, Reuters, Bloomberg, Tom's Hardware, TrendForce, SemiAnalysis, and Chinese trade media, as dated per observation. Spot prices move; treat every figure as of its date, not as a quote.