An interactive model of what it costs to serve one token of a frontier LLM — and the list-price direct-serving contribution margin that implies. Prices and inputs are current as of July 10, 2026 unless a different effective date is shown. Defaults are per-platform anchor fits to public measurements (DeepSeek's serving disclosure, production H20/Ascend deployments, GB200/GB300 benchmarks); every slider documents its sources. Adjust the assumptions — nothing here is any provider's actual ledger.
The claim under examination: Anthropic's marginal cost of serving a Claude token is so far below its API list price that the gross margin on serving is 90–95%. This is a claim about unit serving economics — one token, on a warm GPU, at scale — not about Anthropic's income statement (see §7 for why those diverge).
The loudest and most quantitative proponent is @teortaxesTex:
"math for DeepSeek Serving Opus is at most $4/1Mt. […] They are fleecing you even on subscriptions, except maybe at 100% utilization" — Jun 28, 2026 (elided for slur)
"No, they'll just increase the batch size, have the same speed, and drive margins from 90% to 95%. You're welcome" — Jun 27, 2026
Earlier and more conservative from the same account: "if we exclude R&D and look at inference alone, Anthropic and OpenAI are making like 80% margins" (Mar 2025).
@zephyr_z9 (Zephyr) makes the same claim from the semis side: "At least xAI isn't juicing up the gross margins to 90%-95% and scamming consumers" (Jul 8, 2026) — while separately pegging Anthropic's company-level gross margin at ~70% with 15–20% FCF margin (Jun 24, 2026). Note those two numbers coexisting in one analyst's head: 90–95% on the unit, ~70% on the books.
An attribution check worth flagging: Jukan — a name often cited alongside this claim, identified by the Grok sweep as @jukan05, though GPT Pro's independent browsing instead surfaced posts under @Jukanlosreve (the two engines disagree; both are reported in §6) — was not found asserting 90–95% token margins. His widely-cited threads are about adjacent economics: GPU cluster leasing (xAI's ~220k GPUs at ~$2.60/GPU-hr ≈ $5–6B/yr; "Anthropic converts roughly $5B of spend into what it expects to be $15B of ARR", May 2026) and the memory-bound nature of inference ("Inference is memory", Jun 2026). His famous 90%+ margin posts are about SK hynix DRAM, not Anthropic.
The counter-camp: @fleetingbits — "we know approximately what frontier lab inference margins are; it's like 40-50%; it's been reported a bunch of times. anthropic labels cloud provider commissions as a sales and marketing expense; so the gross margins are mostly inference compute costs." And in the middle, Dylan Patel of SemiAnalysis: Anthropic's margin on an Opus API token is "north of 80%".
On March 1, 2025 (Open Source Week "Day 6"), DeepSeek published actual production serving statistics for V3/R1 on H800 clusters: 73.7k input / 14.8k output tokens per second per 8-GPU H800 node, 56.3% input cache-hit rate, ~$87,072/day GPU cost (at $2/H800-hr) against $562,027/day theoretical revenue at R1 list prices — the famous "545% cost-profit ratio," which is a markup; as a margin it is 84.5% (a distinction TeorTaxes himself insisted on). Caveats DeepSeek listed: much traffic was free web/app users and off-peak V3 pricing, so realized revenue was materially lower than the theoretical figure.
Why it matters for Anthropic: DeepSeek's disclosure implies an 84.5% theoretical margin had all observed traffic been billed at R1 list prices — DeepSeek itself stated actual realized revenue was materially lower — on export-restricted hardware, in early 2025, at prices far below Anthropic's on a like-for-like token-class basis: 3.6× cheaper for cache reads, 9.1× for fresh input, 11.4× for output. The hardware and serving-software inputs to that calculation have since improved (§4), though tariffs, fleet constraints and latency targets can move the other way. The a-fortiori argument — if DeepSeek could get to ~85% theoretical at $2.19/Mtok output on H800s, what does $25/Mtok output on GB300s and TPUv7 imply? — is the strongest single piece of evidence in the bull case. It is an anchor for what optimized serving can cost, not proof of what Anthropic's margin is.
The leak is widely retold as "Musk revealed Opus is much smaller than expected" — but on total parameters it says the opposite. What Musk actually posted (Apr 9, 2026): "0.5T total. Current Grok is half the size of Sonnet and 1/10th the size of Opus. Very strong model for its size." The community deduction (e.g.): Sonnet ≈ 1T, Opus ≈ 5T total parameters — Opus is big.
The "smaller than expected" intuition belongs to active parameters: TeorTaxes, observing Fable served at ~90 tok/s, concluded it had "shockingly FEW active parameters for what it was", and Zephyr puts OpenAI's frontier at "~100B active range" with Opus/Fable "the highest active parameters" among peers. For serving cost, active parameters are what matter (FLOPs/token ≈ 2 × active); total size mostly sets the HBM footprint. A 5T-total/~300B-active MoE is the shape our defaults assume — and the active number is the single most uncertain, most consequential slider on this page. Even the total is contested: the "incompressible knowledge probes" estimation method (arXiv:2604.24827) carries ±3× error bars per model, and a methodological re-analysis lands nearer 1.1T total for Opus 4.7. GPT Pro's sanity check cuts the other way: a dense 5T Opus would cost more than its own $25/Mtok list price to serve — so if Musk's number is right, heavy sparsity isn't optional, it's implied.
The researcher is @_xjdr; the product is ncode on the Noumena platform (code.noumena.com), served on GB300 NVL72 racks from Prime Intellect. His free-week postmortem (Jun 30, 2026) is the best public look at frontier-style serving on Blackwell Ultra:
"final GLM 5.2 served stats: ~12000 unique api keys, served ~300B tokens total, 232 tok/s/gpu output average, 431 tok/s/gpu output max sustained, 2.1 sec TTFT average (1M ctx), 81k tok average input size, 41% cache hit rate"
Configuration: 60× B300 ("15 trays"), bf16 attention with fp8 experts/KV ("virtually 0 measurable quality difference"), no MTP, no Eagle — a custom Rust stack. A third-party estimate on his thread put the implied cost at ~$0.35/Mtok in, ~$1.50/Mtok out at $6/GPU-hr. Note his 232 tok/s/GPU average is output-only bookkeeping on a workload with 81k-token average inputs at 1M context — the GPUs were mostly doing prefill; it is not comparable to SGLang's 12k tok/s/GPU throughput records (different model, batch regime, and MTP).
Adjacent results worth separating from the ncode story (GPT Pro's browsing initially merged them, §6): GLM-5.2's architecture is public via Baseten's serving writeup — 744B total / 40B active, NVFP4-clean, 280+ tok/s/user on Blackwell (a per-user speed record, not aggregate throughput). And the headline GB300 aggregate number belongs to SGLang + NVIDIA serving DeepSeek-V4: 2,200 → 11,200 tok/s/GPU from April to June 2026 on the same racks — five-fold, from software alone.
Covered in §8 — the short version: heavy Claude Max users demonstrably extract 15–40× the subscription price in API-equivalent tokens, which is only economically sane if the marginal cost of those tokens is a small fraction of list price. The plans are themselves evidence for the bull case.
The calculator above prices a token from first principles:
output $/Mtok = ($/GPU-hr ÷ 3600 ÷ tokens/s/GPU) × 10⁶ ÷ utilization, with tokens/s/GPU = (dense FP8 FLOPS × precision factor × effective-MFU × interactivity) ÷ (2 × active params). Prefill (input) uses the same form at higher MFU; cache reads cost a few percent of prefill. The workload mix (input:output ratio, cache-hit rate) blends the three, and the same mix prices revenue at list minus cache/batch/negotiated discounts.
Anchor fits, not vibes — and not a validated predictive model. Each accelerator's effective-MFU is fitted so the model reproduces that platform's best public measurement. That is six separate anchor fits, not one calibrated theory: we ran the obvious falsification test — fit a single global MFU on the DeepSeek anchor alone and predict the other platforms — and it fails (mean error 37%, worst 59%; full write-up in the LOAO methods note). Consequences we adopt: reproducing an anchor is an identity, not a validation; per-platform values are interpolation near their anchored operating points; and platforms with no published anchor (TPU v7, Trainium, the Rubin projection) carry materially lower confidence than the anchored rows. Validation anchors are also not always the neutral defaults — where a published anchor comes from a vendor-optimized or latency-relaxed configuration, the deployed default is the hardware dive's neutral recommendation, and the table shows both:
| Anchor (published) | Published | Reproduces at | Deployed default |
|---|---|---|---|
| DeepSeek disclosure, H800 decode (37B active, FP8, production average) | 1,850 tok/s/GPU | 1,873 at MFU 7.0% | 7.0% |
| DeepSeek disclosure, H800 input flow — includes the 56.3% disk-cache-hit share, so it is not a fresh-prefill benchmark | 9,212 tok/s/GPU aggregate | not used directly | fresh prefill 15% ≈ 4,013 tok/s (cache-share reconstruction ≈ 4,026; v1 wrongly used 34%) |
| vLLM GB200, R1 decode (source precision basis not fully pinned — possibly NVFP4) | ~10,100 tok/s/GPU | 10,135 at MFU 15% — using GB200's verified 5.0 PF dense per GPU (an earlier revision missed by −10% by using B200's 4.5 PF, a sparse/dense mixup common in secondary sources) | 15%, treated as an upper anchor |
| SGLang GB300 record, V4 Pro 1.6T FP4 + MTP (49B active, disclosed) | >12,000 tok/s/GPU | ~12,050 at 12.7% × 1.85 FP4 (refit after the 49B disclosure; v1 fit assumed ~66B active) | 12.7% |
| Ant Group/SGLang production, H20 decode (R1, FP8, relaxed <70 ms tier) | 714 tok/s/GPU (675 at <50 ms; 423 at <30 ms) | 714 at MFU 18% | 17% neutral ≈ 680 tok/s (the tighter tier) |
| CloudMatrix-Infer, Ascend 910C decode (R1, INT8; vendor-measured, optimized 384-NPU supernode) | 1,943 tok/s/NPU | 1,943 at MFU 9.5% of 1.504 PF INT8 | 7% neutral ≈ 1,422 tok/s (DeepSeek's "60% of H100" eval implies 5.5–6.5%) |
Those single-digit decode MFUs are not a bug — decode is memory/interconnect-bound; this is why "inference is memory" and why bigger NVLink domains and HBM keep beating raw FLOPS. The H20's seemingly heroic 17–18% is the same physics from the other side: a bandwidth-rich chip whose FLOPS denominator is tiny — and it is exactly why a single MFU cannot transfer across platforms. Domain of validity: short-to-moderate context, throughput-oriented serving; context length, TTFT/TPOT targets and KV-cache lifecycle are not modeled, and the latency curve is steep (CloudMatrix drops 1,943 → 538 tok/s when TPOT tightens from ~50 ms to 15 ms).
| Platform | Dense FP8 (PF) | Dense FP4 (PF) | HBM | BW | TDP | Rental (Jul 2026) |
|---|---|---|---|---|---|---|
| H100 SXM (2022) | 1.98 | — | 80 GB | 3.35 TB/s | 700 W | $1.99–3.90/hr, spot ≈ $2.40 |
| H200 (2024) | 1.98 | — | 141 GB | 4.8 TB/s | 700 W | $2.45–4.50/hr |
| B200 HGX, per GPU (2024) | 4.5 | 9 | 180 GB | 8 TB/s | 1.0 kW | $6.69/hr on-demand (Lambda) |
| GB200 NVL72, per GPU (2024-25) | 5.0 | 10 | 186 GB | 8 TB/s | 1.2 kW | $3.50–6.00/hr; rack ≈ $3–3.5M |
| GB300 NVL72 "Blackwell Ultra" (2025-26) | 5.0 | 15 | 288 GB | 8 TB/s | 1.4 kW | $4–7/hr early; rack ≈ $3.5–4.5M |
| TPU v7 Ironwood (GA Apr 2026) | 4.61 | — | 192 GB | 7.37 TB/s | ~1 kW | undisclosed; Anthropic deal: up to 1M chips |
| Trainium2 (GA Dec 2024) / Trainium3 (GA Dec 2025) | 1.3 / 2.51 | — | 96 / 144 GB | 2.9 / 4.9 TB/s | ~0.5–0.8 kW | AWS-internal; Rainier launched with ~500k Trainium2; Anthropic reported >1M in use by Apr 2026 |
| H800 — China export SKU (2023) | 1.98 | — | 80 GB | 3.35 TB/s | 700 W | IDC annual-commit $1.47–2.06/hr (mid-2026, +30% post-Spring-Festival); DeepSeek's disclosure assumed $2 |
| H20 — China-legal SKU (2024) | 0.296 | — | 96 GB | 4.0 TB/s | 400 W | ~$10–12k chip / ~$20k installed; rental class spans ~$0.76 (IDC annual) to $7+ (cloud on-demand) |
| Huawei Ascend 910C (2024–25) | 1.50 (INT8; no FP8) | — | 128 GB | 3.2 TB/s | ~0.6 kW | ~$23k/chip installed; Huatai procurement $1.71–2.25/hr; CloudMatrix 384 ≈ RMB 60M |
| Vera Rubin NVL72, per GPU (preliminary 2026 specs) | 17.5 (dense FP8/FP6) | 50 sparse NVFP4 inference (35 dense-class training) | 288 GB HBM4 | 22 TB/s | ~1.8 kW | production power, throughput and cost TBD |
Measured, not marketing: SemiAnalysis InferenceX (the successor to InferenceMAX, the field's independent benchmark) finds the most-optimized GB300 NVL72 delivers ~17× the best H100 config in FP8 and ~32× in FP4 on DeepSeek R1 (Jun 27, 2026) — and, critically, that software alone was a 14× gain on the same silicon (baseline FP8 ~1k → wideEP+disagg ~8k → +MTP ~14k tok/s/GPU). SGLang's production numbers agree: 12k tok/s/GPU on V4 Pro 1.6T, 6.5× over B200 with Dynamo disaggregation. At rack level: an H100 rack two years ago did ~8.8k tok/s; a GB300 NVL72 does ~370k — 42× in two years (8× of it HBM growth).
The China stack is its own cost universe. Under export controls the Chinese labs serve on three tiers: hoarded pre-ban H800s (H100 compute, capped NVLink — where DeepSeek's disclosure happened), the deliberately compute-starved but bandwidth-rich H20 (the main legal SKU since mid-2025; decode-friendly because decode is bandwidth-bound — Ant Group's production SGLang deployment is the best public anchor), and Huawei's Ascend 910C (Huawei's CloudMatrix-Infer paper reports 1,943 tok/s/NPU decode on R1, while DeepSeek's internal evaluation put the chip at ~60% of H100 — vendor and customer numbers disagree, so the calculator's Ascend row carries wide error bars). H200s were license-cleared for ten named Chinese firms in early 2026 but essentially none had shipped as of May; Beijing reportedly moved to allow those purchases this week (Jul 7–8). Chip scarcity also cuts the other way on price: Chinese H-series lease rates rose 20–30% over early 2026 on a >1,000× surge in national token volume (TrendForce) — Chinese margins are being squeezed from the cost side at exactly the moment Western $/token falls. One more trap: "the China rental rate" is a category error — the same H20 spans roughly ¥5–72 per card-hour (>10×) between annual IDC bare-metal leases and hyperscaler on-demand instances. The calculator's China rows use annual-commit rates; the GPU-hour cost multiplier is the dial for other rental classes.
But cost per token falls slower than throughput rises, because rental prices track capability: $2.40/hr H100 → $6/hr GB300 eats ~2.5× of the 17×. Net hardware-economics gain per generation in $/Mtok: H100→H200 ~1.2×, H200→GB200 ~2–3×, GB200→GB300 ~1.1–1.5× (GB300's edge is HBM capacity for reasoning/long-context, not FLOPS), Rubin ~2–3× again. Cumulative 2024→2027: roughly 6–25× cheaper per token at the hardware level, before model-side efficiency (sparser MoEs, MTP, quantization) which historically contributed as much again. This is why margins at constant list prices mechanically march toward 95%+ — and why in practice prices fall instead (Opus's 2025-11 cut to $5/$25 was 3×; fast mode — $10/$50 on Opus 4.8, down from 4.7's $30/$150 tier retiring July 24 — shows the latency premium being monetized separately, and itself falling).
Verdict: conditional, not blanket. A 90–95% list-price serving contribution margin is plausible under strategic-partner or owned-TCO compute, optimized batch-heavy workloads and high fleet occupancy — and 95% is not the ceiling for output tokens on the newest hardware. Under this page's own open-market rental scenario at 50% utilization, the Opus-class result is materially lower: ~75–88% blended, with the deployed calculator's default landing at ~77%. The claim's truth is a function of the invoice and the operating point, and it sits atop a stack of margins that shrinks at every step toward the income statement.
My point estimates for mid-2026, at list prices, balanced latency, 50% fleet utilization, neocloud-level hardware costs:
Three things the bull case gets right: (1) the DeepSeek a-fortiori argument is sound and the only primary-source-grounded argument anyone has; (2) caching is a margin machine — cache reads bill at 10% of input price but cost ~1–5% of prefill to serve, so agentic traffic stays high-margin even at deep effective discounts; (3) each hardware generation adds margin at constant prices, and Anthropic sits on three platforms (GB-class NVIDIA, up to 1M TPU v7, ~500k Trainium) with pricing leverage none of its open-weight competitors have.
Three things it elides: (1) utilization — fleets are provisioned for peak; at 35% utilization my Opus median drops ~8 points; (2) the latency premium is not free — interactive serving at user-acceptable speed costs ~1.4–3× throughput-optimal serving (fast mode's price premium exists for a reason — 2× on Opus 4.8, and it launched at 6×); (3) subscriptions and whales — documented tail users extract 15–40× their subscription in list-value tokens (§8); how much of the subscriber base is underwater is unknowable without a usage distribution nobody publishes, but the leakage direction is real and the API numbers never show it.
Forward: on GB300→Rubin economics, at unchanged list prices, Opus-class output margins pass 92–96% by 2027; I expect Anthropic to keep converting the surplus into price cuts and reasoning-token volume instead of letting the sticker margin print, so the quoted marginal margin stays in the high-80s/low-90s while $/token falls another 3–5×. The interesting question isn't whether 90–95% is right — it's how long the industry can hold list prices this far above a falling cost floor with open-weight models (DeepSeek V4, GLM 5.2, Kimi) selling adequate quality at 10–20× less.
GPT-5.6 Pro's verdict: "Standard-list-price marginal serving margin: approximately 92–94% for Opus and 94–96% for Sonnet on a mature 2026 fleet." And: "Is 95% an upper bound? No." — Sonnet at normal $3/$15 pricing exceeds it (96.3% on output), Opus on strategic TPU contracts exceeds it (97.2%), and the electricity-only marginal margin is >99.8%. But 95% is "not conservative for every token": at public cloud rates the same math yields only 55–85%.
Method. It priced an "economic marginal cost" — accelerator opportunity cost at Anthropic-scale contract rates, 75% fleet utilization, +10% CPU/network/reliability overhead — explicitly not electricity-only and not full COGS. Architecture: Opus ≈ 5T total / 300B active (independently landing on the same activation ratio as §5, ~6%, matching DeepSeek's 5.5% and GLM-5.2's 5.4%), FLOPs/token = 2.3 × active. Fleet blend: 40% TPU v7, 25% GB300, 15% GB200, 15% Trainium2, 5% H200.
Central estimates (Opus, $/Mtok serving cost, input → output):
| Basis | Input cost → margin @ $5 | Output cost → margin @ $25 |
|---|---|---|
| Fleet blend (central) | $0.47 → 90.6% | $1.68 → 93.3% |
| GB300 NVL72 only | $0.31 → 93.8% | $1.44 → 94.2% |
| TPU v7 at est. strategic rate (~$1.60/hr per SemiAnalysis) | $0.24 → 95.1% | $0.70 → 97.2% |
| H100/H200 (economic rates) | ~$1.2 → ~76% | ~$4.9 → ~80% |
| Public cloud rates (AWS H100 $5.19/hr … B300 $14.04/hr) | margins collapse to 55–85% — "why both sides of the online argument can produce superficially plausible calculations" | |
Sensitivity it flagged as dominant: active parameters. Opus at 150B active → 95–97% margins; at 600B active → 81–87%. Identical conclusion to §5: the whole debate compresses into one undisclosed number.
New evidence it surfaced (not in the X sweep): PitchBook/Morningstar's June 2026 estimate of Anthropic gross margin ≈ 44%, with compute spend of $0.71 per revenue dollar in Q1 falling to a projected $0.56 in Q2 — the freshest business-margin datapoint anyone has; GLM-5.2's actual architecture (744B total / 40B active) via Baseten's "world's fastest GLM-5.2 API" writeup (280+ tok/s/user on Blackwell NVFP4); and the SGLang+NVIDIA GB300 progression 2,200 → 11,200 tok/s/GPU between April and June 2026 — a 5× gain from software alone on fixed hardware, its exhibit for why "serving software is a first-order economic variable." Its illustrative bridge from 93.5% marginal margin down to the ~44% reported gross margin allocates: −12 pts price realization, −8 peak provisioning, −7.5 cloud markup/older hardware, −9 free traffic + subscription whales, −13 CPU/network/retries/safety/internal inference.
Attribution divergence, preserved: GPT Pro's browsing never located the ncode/Noumena deployment and instead attributed the GB300 GLM-5.2 story to a merger of the SGLang/NVIDIA DeepSeek-V4 result with Baseten's GLM-5.2 work. The Grok X sweep, however, found the actual account — @_xjdr's ncode/Noumena, with primary post URLs (§2c) — while GPT Pro's browsing surfaced hardware-cost posts under the handle @Jukanlosreve, contradicting Grok's @jukan05 identification. Where the two engines' account attributions conflict we show both; the direct post links in §1–2 are the ground truth we could verify.
Where GPT Pro and I (Claude) differ — and why it's smaller than it looks: its central Opus numbers run ~10 points above my rental-based median (93.3% vs ~80% blended). The gap is almost entirely procurement assumptions, not physics: it prices hardware at Anthropic's negotiated strategic rates (TPU ~$1.60/hr) and 75% utilization; my median uses open-market neocloud rental at 50% utilization. Run my model in owned-TCO mode at its utilization and the two converge at ~93% — which is exactly what the "SemiAnalysis / Dylan Patel" and "GPT-5.6 Pro fleet model" presets above let you reproduce. The honest summary: the serving margin at list price is roughly 78–85% if you think Anthropic pays market rates for compute, roughly 90–95% if you think it pays hyperscaler-partner rates. Procurement explains most of the divergence between analysts on the same model — while architecture (active parameters above all), workload, latency targets and utilization remain first-order unknowns in their own right.
A rare piece of direct evidence on the invoice question surfaced in the xAI dive (§10): the SpaceXAI prospectus discloses that Anthropic pays xAI $1.25B per month for ~325,000 GPUs plus supporting CPUs, storage and networking — about $5.27 per bundled GPU-hour, a real, dated, arm's-length price for capacity Anthropic actually buys. Caveats: it is a bundled service price, not a bare chip-hour, and contracted reserve capacity need not price the marginal fleet. But it brackets the debate from above: Anthropic's blended compute invoice cannot plausibly sit far above a rate it demonstrably pays a third party, and that rate sits between this page's "market rental" and "strategic partner" scenarios — closer to the former.
Reported/leaked figures for Anthropic's business-level gross margin paint a different picture from the unit math. The Information (Jan 2026, from people with knowledge of its financials): Anthropic lowered its 2025 gross-margin projection to 40% (down 10 points from the earlier 50% plan) because inference costs on Google/Amazon servers ran 23% higher than anticipated; including free-tier chatbot inference the margin would be ≈38%. Same reporting: 2025 revenue ≈ $4.5B (~12× 2024's $381M), 86% of it API. SemiAnalysis put the 2024 accounting gross margin at −94% (yes, negative). Zephyr's mid-2026 read is ~70% GM with 15–20% FCF margin; a detailed July 2026 thread claims quarterly gross profit swung from −$55M to ~$453M with inference cost/token down ~40× since early 2024 (@IvanaSpear). The freshest independent estimate is PitchBook/Morningstar (Jun 2026): gross margin ≈ 44%, with compute spend of $0.71 per revenue dollar in Q1 2026, projected $0.56 in Q2. The trajectory — deeply negative (2024) → ~40% planned (2025) → ~44% estimated (mid-2026), with bulls like Zephyr already seeing 70% — is the margin ladder being climbed in public, and The Information ran a follow-up on why the labs kept missing their own gross-margin forecasts.
The bridge from ~90% unit margins down to ~40–70% books, roughly in order of size. Illustrative and non-additive: these buckets overlap and are not an accounting reconciliation — they show how alternative procurement, occupancy and workload assumptions move the modeled result:
| Bridge item | Mechanism | Rough magnitude |
|---|---|---|
| Cloud markup | Anthropic buys most compute from AWS/GCP, who take their own margin; some channel revenue also shares 30–40% with the clouds (Bedrock/Vertex economics). DeepSeek/High-Flyer appear to control substantial compute (though the $2/hr figure was a costing assumption and the current owned/rented mix is undisclosed); Anthropic primarily consumes partner-provided capacity, whose long-term committed economics can sit well below public cloud rental. | −10 to −20 pts |
| Utilization & peak provisioning | Capacity sized for Monday-morning peak; the unit math assumes a busy GPU. 35% vs 70% utilization is a 2× on cost. | −5 to −15 pts |
| Subscription over-consumption | Flat-fee plans where the tail extracts 15–40× the fee (§8); weekly caps (Aug 2025) exist to clamp exactly this. | −5 to −10 pts |
| Free tier & internal inference | claude.ai free traffic, evals, RL/synthetic-data generation all burn serving compute against zero revenue (some booked as R&D, treatment varies). | −5 to −10 pts |
| Discounts & mix | Enterprise/committed-use discounts, 50%-off batch tier, cache-read-heavy agentic mix pulling realized $/Mtok toward $0.99 observed vs $5/$25 sticker. | −3 to −8 pts |
| Accounting choices | What lands in COGS vs S&M vs R&D differs by lab — fleetingbits' point that Anthropic books cloud commissions under sales & marketing cuts the other way, flattering GM. | ± |
Training compute — the industry's historically heavy cash burn — sits below gross margin in R&D — it explains why the company loses money overall (training reportedly falling from 400%+ of revenue toward ~36% by end-2026 per the IvanaSpear thread), not why gross margin is below the unit margin.
The best single investigation found: ksred's Claude Code pricing guide — longitudinal tracking of ~10B tokens over 8 months: ~$15,000 at API list prices against ~$800 of Max subscription fees (~19×). The canonical X-thread calculation is @melvynx (Jan 2026): $50 of tracked usage = 6% of a week's Max-20x allowance ⇒ ~$800/week ⇒ ~$3,200/month of API-equivalent capacity for $200. Corroborating longitudinal data points: 14B+ tokens / $7.5k ccusage in 2.5 months, $8,030 in 4 weeks, 8.6B tokens / ~$8.5k over 60 days (~21× leverage), and the pathological single-session case: $4,243 at token rates in one session on a $100 plan.
What these investigations do and do not establish: heavy subscribers demonstrably extract API-list-value equivalents many times their fee — $3,200 of sticker sold for $200 in the melvynx case, which at this page's modeled direct serving cost is perhaps $300–900 to actually serve depending on mix. They characterize the tail, not the median — no representative usage distribution, mean or median is public, so nothing here identifies whether the typical subscriber is profitable. What the plans' generosity is consistent with is low marginal serving cost; it does not independently prove it — breakage, throttling, workload routing, acquisition subsidy and cross-subsidy can all contribute. The weekly limits (Aug 2025) and repeated 5-hour-window retuning (May 2026 doubling) show the tail is real enough to clamp. The subscription card in the calculator computes break-even for whatever usage level you set — it says nothing about the distribution.
The Anthropic verdict above rests on an unusual density of evidence: a primary serving disclosure from a direct competitor, a parameter leak from a rival CEO, leaked business financials, and a live X-sphere argument between named analysts. None of the other providers has all of that, and some have almost none of it. Read the cards below as provider-native case studies, not a like-for-like ranking: each headline reports the metric its underlying deep dive estimated — OpenAI, Google, xAI, DeepSeek and Zhipu are workload-specific blended list-price serving contribution margins, while Moonshot is an output-token margin only — and the workload mixes, service tiers and cost lenses differ materially across cards, so rank order would encode the analyst's assumptions as much as the providers' economics. The displayed ranges are judgmental uncertainty ranges intended to contain roughly 80% of plausible modeled outcomes — not statistically calibrated confidence intervals. The badge grades public observability of the evidence base, not confidence in the number, and each card carries a five-dimension evidence profile inside. Every headline is reproducible in the calculator above: select the model and the "§10 dive replay" perspective — a deterministic test suite holds each within a point.
Evidence profile — architecture: medium · pricing: high · fleet & TCO: medium · production throughput: low · financial perimeter: medium
The closest analogue to the Anthropic case: a partner-hosted fleet (OpenAI discloses 3 GW of dedicated inference capacity on Hopper/Blackwell across Microsoft, OCI and CoreWeave — contractually controlled but not owned), premium list prices (Sol $5/$30 short-context), and an active-parameter estimate (~100B) that is only a community figure — though one independently corroborated by Epoch's inference-economics work. The most interesting tension: a ~94% modeled list-price margin coexists with a reported 70% "compute margin for paying users" (The Information, Oct 2025) and a 33% company adjusted gross margin — the gap is take-or-pay reserved capacity, subscriptions and free traffic, i.e. the same margin-stack that separates Anthropic's unit math from its books (§7).
Why the interval is 86–97%: throughput assumptions alone span >4× in cost; reserved take-or-pay capacity can make slack-period allocated cost several times the engineering marginal cost; and the true transfer price is bracketed only by market comparables and Oracle-offtake arithmetic.
Evidence profile — architecture: low · pricing: high · fleet & TCO: medium · production throughput: low · financial perimeter: medium
The highest central margin on the page, for a structural reason: Google is the only provider serving a frontier model on silicon it designs, in datacenters it owns, at internal cost — a derived ≈$1.28/Ironwood-hour (vs the $12 on-demand public rate, and below even Anthropic's reported ~$1.60 strategic rate). But it carries the weakest architecture evidence of any provider: no credible leak of Gemini's total or active parameters exists at all, so the 120B-active/3T-total inputs are scenario midpoints, not estimates. The disclosed facts are all about scale and trajectory: serving unit costs down 78% during 2025, ~3.2 quadrillion tokens/month across surfaces, ~19B API tokens/minute.
Why the interval is 89–98%: the two dominant unknowns (active params, effective decode MFU) each span 3–4× in cost, but vertical integration puts a floor under the answer — a sub-90% result lies toward the costly end of the stated architecture, throughput and occupancy ranges — public evidence does not rule it out.
Evidence profile — architecture: medium · pricing: high · fleet & TCO: high · production throughput: low · financial perimeter: high
The one lab that owns (or at least operationally controls) its fleet — and the widest confidence interval on the page, because "what does a GPU-hour cost xAI?" has three defensible answers. The SpaceXAI prospectus discloses a fleet of >440k accelerators (200k+ H100/H200/GB200 at Colossus, 110k GB200 + 110k GB300 at Colossus II) — but $20.2B of it sits on finance leases, and AI capex ran $12.7B in 2025 alone. Grok 4.5 (launched Jul 8 at $2/$6) is disclosed at 1.5T total parameters; active count is anyone's guess (100–500B). The decisive fact: Anthropic pays xAI $1.25B/month for ~325k GPUs (≈$5.27/GPU-hr) — so every GPU-hour spent serving Grok forgoes a disclosed wholesale price more than twice xAI's estimated full-cycle cost. At cash-marginal cost Grok's margin is ~92%; at full-cycle TCO ~67%; at the Anthropic-contract opportunity cost, serving output-heavy traffic barely beats selling the capacity. Zephyr's "xAI isn't juicing margins" is true or false depending entirely on which lens you pick.
Why the interval is 10–85%: unknown batch throughput spans ~8× in cost, and the GPU-hour can be honestly valued anywhere from $0.60 (cash) through $2.40 (full-cycle) to $5.27 (contracted opportunity cost) — the margin question dissolves into the valuation question.
Evidence profile — architecture: high · pricing: high · fleet & TCO: low · production throughput: low · financial perimeter: low
The inversion of the Anthropic story. DeepSeek has the best evidence of any provider — disclosed architecture (1.6T total / 49B active, selective FP4), disclosed pricing, and the field's only production serving disclosure — yet one of the lower provider-native estimates (the cards are not directly rankable), because it chose to convert its efficiency into price: a permanent 75% cut took V4 Pro to $0.435/$0.87, roughly 60% below R1's old output price. Repricing the old disclosed workload at today's tariff turns the famous 84.5% margin into ~67% before any cost improvement — and V4-Flash would be underwater on the old cost structure. The announced mid-July 2× peak-hour surcharge would restore ~84.5% in peak windows: the old margin, now sold as a surge price. Export controls cut both ways on the cost side: they constrain hardware choice and add software/fleet-fragmentation overhead (the fleet is now reportedly part Huawei Ascend 950), yet current H800 asking rates sit below the old $2/hr basis — the net effect cannot be signed from public evidence.
Why the interval is 45–83%: the estimate scales the disclosed 2025 baseline by two unknown factors — hardware-hour cost (0.5–1.1× the old basis) and per-token work for the bigger-but-sparser V4 (1.05–1.5×) — which compound to a 3× cost span.
Evidence profile — architecture: high · pricing: high · fleet & TCO: low · production throughput: low · financial perimeter: high
The rare provider with audited numbers: Zhipu's Hong Kong listing (Jan 2026) forces disclosure nobody else makes. Its cloud/API segment ran a −0.4% gross margin in H1 2025 (price-war casualty) recovering to 18.9% for FY2025 — and with API prices up 83% since end-2025, a price-only counterfactual lands in the mid-50s, which is what the marginal estimate here reflects. The architecture is fully disclosed (744B/40B, open weights), and third-party resellers bound the cost from above: DeepInfra sells GLM-5.2 below Z.ai's own list. The wrinkle: Zhipu’s filings show sharply lower capex (down 84% in 2025) and a shift toward one-to-four-year compute-service contracts — the current owned-vs-rented mix is undisclosed and says it serves across nine domestic chip platforms, so its cost basis is contract-opaque and its Coding Plan (~242k paying developers) can be deeply subsidized for heavy users.
Why the interval is 35–77%: decode throughput and the confidential compute price each move the answer ±10–20 points; the audited 18.9% blended floor and the reseller price ceiling are what keep the interval from being wider still.
Evidence profile — architecture: high · pricing: high · fleet & TCO: low · production throughput: medium · financial perimeter: low
The highest raw headline among the three Chinese-provider cards — but it is an output-token margin, not blended, so it cannot be ranked against DeepSeek’s and Zhipu’s blended figures. The economics are legible though: K2.7 kept a premium output price ($4.00/M — 4.6× DeepSeek V4 Pro's post-cut $0.87) on an aggressively cheap-to-serve architecture — 1T total but only 32B active, in natively quantized selective INT4 that fits an 8-GPU replica. The open weights make the cost side unusually boundable: a reproducible 128×H200 benchmark puts the short-context decode-accelerator floor at ~$0.21/M output, SemiAnalysis's same-lineage runs span $0.14–$1.00/M by hardware and speed, and DeepInfra resells K2.7 at $3.50/M — a weak market-price ceiling (its contribution margin is undisclosed, so this is conditional evidence, not a cost bound). Business context: ARR reportedly passed $300M by mid-June with API revenue above 70% — Moonshot is the most API-dependent of the Chinese labs, which is exactly where these margins live.
Why the interval is 55–91%: the open weights pin the architecture but not the deployment — per-user speed targets and the actual fleet each move serving cost 2–3×; the $0.21 decode floor and the $3.50 reseller price bracket the answer from both sides. Note this is an output-token margin; a blended margin needs Moonshot's undisclosed traffic mix.
For readers who want one comparable row per provider anyway, this table is computed live by the calculator with the lens held fixed (open-market rental, 50% utilization, balanced latency — the "evidence median") and the workload held fixed (15:1 input:output, 60% cache hits), while keeping each provider's own prices, cache tariff and fleet. This is deliberately not each provider's operating point — it prices everyone as if they procured compute the same way — and DeepSeek V4's negative number is the honest consequence of its post-price-war tariff under Western rental economics (its own operating point is the ~69% card above). Margins are rounded to whole points; the §10 ranges still apply.
Method note: X posts located and quoted via a Grok 4.5 agent sweep (2026-07-09); hardware/TCO data via Parallel deep research + Exa across SemiAnalysis, MLCommons, NVIDIA/Google/AWS primary pages; cross-model verification via an independent GPT-5.6 Pro research run. The per-provider audits in §10 come from one dedicated GPT-5.6 Pro deep dive per provider plus two Chinese-accelerator hardware sweeps (all 2026-07-09, archived unedited in the research annex). Attributions (Jukan's handle, the direction of the Musk leak, the ncode/Noumena identification) were verified against primary posts; where the research engines disagreed, both readings are shown rather than harmonized. Model sizes for closed models remain estimates — the calculator exists so you can disagree with a slider instead of a screenshot.