← Frontier Inference Margins · all research reports
Run 2026-07-09 · 53m24s · model
gpt-5-6-pro(Pro-verified, Ashitaorbis project) · conversation Verdict: Grok 4.5 marginal serving gross margin ≈ 67% central, 80% CI ≈ 10–85% at API list price under full-cycle owned TCO; ~92% under strict cash-marginal cost; ~29% at the Anthropic-contract opportunity cost. Zephyr's "not juicing margins" claim is definition-dependent. Archived verbatim below (DOM-extracted; math notation may render with minor duplication artifacts).
Snapshot through close of July 9, 2026. Labels mean: DISCLOSED = xAI/SpaceXAI, an executive statement, or a regulatory filing; CREDIBLY REPORTED = reputable reporting based on internal documents or identified sources; COMMUNITY ESTIMATE = named independent public analysis; SPECULATION = my inference, assumption, or arithmetic model. Arithmetic from disclosed inputs is labeled SPECULATION — derived from DISCLOSED inputs.
| Role | Model | Total parameters | Active parameters per token | Architecture | Likely serving precision |
|---|---|---|---|---|---|
| Flagship | Grok 4.5 | 1.5T — DISCLOSED by Musk as the "1.5T V9 foundation model" underlying 4.5 | ~200B central; 100–500B 80% range — SPECULATION | MoE — DISCLOSED by co-developer Cursor; expert count, routing width undisclosed | FP8 or another 8-bit weight format; BF16/FP8 activations — SPECULATION |
| Workhorse | Grok 4.3 / Grok 4.20 family | 0.5T — DISCLOSED by Musk for "current Grok" on April 9; mapping to 4.20/4.3 medium confidence | ~100B central; 80–220B 80% range — SPECULATION | Probably MoE — SPECULATION | FP8 weights, BF16/FP8 activations — SPECULATION |
Musk said Grok 4.5 was based on xAI's "1.5T V9 foundation model" — DISCLOSED. Cursor, which jointly trained the model, described Grok 4.5 as a mixture-of-experts model — DISCLOSED. Musk's statement · Cursor's technical launch post
The ~200B active central — SPECULATION ≈ 13% of total. The 100–500B range is broad because neither expert count nor experts-per-token is public. Historical prior: official Grok-2 config — 8 routed experts, 2 selected, shared MoE block, 64 layers, BF16 — DISCLOSED; a community count against official weights finds 269.5B total / 115.0B active = 42.7% active — COMMUNITY ESTIMATE. Applying that ratio to 4.5 would give ~640B active — plausible upper tail, too dense for the central. Grok-2 config · community count
xAI says Grok 4.5 serves at 80 output tokens/s — DISCLOSED (user-stream, not replica throughput), 500k context — DISCLOSED, trained on tens of thousands of GB300s — DISCLOSED. Launch post
Musk's April 9 statement: 0.5T total — DISCLOSED for "current Grok" (half of Sonnet, one-tenth of Opus). Current workhorse endpoints have 1M context — DISCLOSED. Musk · pricing
No DISCLOSED production precision. Official Grok-2 serving recipe: FP8, TP8 — DISCLOSED (best prior). 1.5T weights = 0.75TB at 4-bit / 1.5TB at 8-bit before KV etc. Central case: 8-GB300 replica (6–16 range) — SPECULATION.
| Site / phase | Mid-2026 inventory | Status |
|---|---|---|
| Colossus/C1 initial | ~100,000 H100 | DISCLOSED (prospectus) |
| C1 after expansion | >220,000 GPUs (H100/H200/GB200) | DISCLOSED |
| H100 within current C1 | 200,000 H100 | DISCLOSED (Colossus page) |
| Colossus II cluster 1 | ~110,000 GB200 (210MW) | DISCLOSED |
| Colossus II cluster 2 | ~110,000 GB300 (220MW) | DISCLOSED |
| Next C2 phase | ≥220,000 more GB300 | DISCLOSED as planned (excluded) |
| Conservative installed floor | >440,000 accelerators | SPECULATION — derived from DISCLOSED |
SpaceXAI prospectus pp. 63–64 · Colossus page · Anthropic compute announcement
Three related-party Valor equipment leases carry aggregate undiscounted payments of $6.986B + $6.633B + $6.587B = $20.206B — DISCLOSED / derived. "Owned fleet" = controlled dedicated infrastructure, not unencumbered title.
| Cost lens | Central | 80% range | Interpretation |
|---|---|---|---|
| Strict short-run cash marginal | $0.60/hr | $0.30–0.90 | Power, cooling, maintenance, ops after sunk hardware |
| Accounting serving TCO | $1.8/hr | $1.3–2.4 | + straight-line depreciation |
| Economic full-cycle owned TCO | $2.4/hr | $1.8–3.2 | Replacement capital, financing, obsolescence — central verdict input |
| External opportunity value | $5.27/hr | — | The Anthropic contract price |
| Future Google contract value | $11.45/hr | — | From October 2026 |
Capital anchor: AI capex $5.633B (2024) + $12.727B (2025) + $7.723B (Q1 2026) = $26.083B — DISCLOSED/derived; ~$65k all-in installed per accelerator-equivalent (SPECULATION), 5.5-yr server life, 10% capital charge, 85% allocatable hours → ~$1.8/hr capital + $0.6/hr cash ops = $2.4/hr.
Power: enumerated phases imply 1.75kW nameplate per accelerator — derived; prospectus reports 1.0GW nameplate compute draw as of Mar 31, 2026 — DISCLOSED. Memphis 2026 GSA tariff: ~$0.060–0.065/kWh energy + $16.51–18.19/kW-mo demand — DISCLOSED → ~$0.22/hr grid power. MLGW tariff
Jukan's ~220k C1 GPUs at $2.60/hr ≈ $5.01B/yr — COMMUNITY ESTIMATE vs my full-cycle TCO $4.63B/yr — ownership advantage only ~8% at full cycle. The dramatic advantage exists only in the short-run cash view ($1.16B/yr).
The market comparison that matters: Anthropic pays $1.25B/month for ~325,000 GPUs + CPUs/storage/networking — DISCLOSED = $5.27/GPU-hr — derived. Google (from Oct 2026): $920M/month for ~110,000 GPUs — DISCLOSED = $11.45/GPU-hr — derived. These are reserved-capacity sale prices, not production cost — but they price xAI's opportunity cost. Prospectus · Google agreement, SEC
Internal memo reportedly put training MFU at ~11% ("embarrassingly low"), target 50% — CREDIBLY REPORTED (Business Insider) vs 35–45% typical. Training-stack evidence only — not inference occupancy. Commercial evidence of capacity monetization: 325k GPUs to Anthropic, tens of thousands to Cursor training. Business Insider
| Endpoint | Context | Input | Cached input | Output |
|---|---|---|---|---|
grok-build-0.1 |
256k | $1.00 | $0.20 | $2.00 |
grok-4.5 |
500k | $2.00 | $0.50 | $6.00 |
grok-4.3 |
1M | $1.25 | $0.20 | $2.50 |
grok-4.20-0309-reasoning |
1M | $1.25 | $0.20 | $2.50 |
grok-4.20-0309-non-reasoning |
1M | $1.25 | $0.20 | $2.50 |
grok-4.20-multi-agent-0309 |
1M | $1.25 | $0.20 | $2.50 |
Pricing. Reasoning tokens bill at normal rate — DISCLOSED. Grok 4.3/4.20 get a 20% batch discount — DISCLOSED; Grok 4.5 has no batch discount. Priority = 2×. Cursor sells a channel-specific "fast" 4.5 at $4/$18 — DISCLOSED by Cursor. Cursor
Grok Free $0; SuperGrok $30/mo — DISCLOSED; SuperGrok Lite ~$10 (CREDIBLY REPORTED); SuperGrok Heavy ~$300 (CREDIBLY REPORTED); X Premium+ $40/mo or $395/yr — DISCLOSED. xAI pricing · X Premium
117M MAU of Grok AI features (Mar 31, 2026) — DISCLOSED, up from 89M at year-end 2025; no paid-subscriber or token-volume disclosure ⇒ no defensible realized $/token. A $30 SuperGrok = only 5M Grok-4.5 output tokens at list — heavy users realize far below list.
| Evidence | Number | Label | Interpretation |
|---|---|---|---|
| Q1 2026 AI-segment revenue | $818M | DISCLOSED | X ads, subscriptions, licensing, AI products, compute infra — not Grok API alone |
| Q1 AI cost of revenue | $456M | DISCLOSED | Infra, energy, bandwidth, depreciation, cloud, rev shares, ops |
| Implied Q1 AI gross margin | 44.3% | derived | Segment accounting margin, not marginal serving margin |
| Q1 AI R&D | $2.379B | DISCLOSED | Includes training infra + GPU depreciation |
| Q1 AI operating loss | $2.469B | DISCLOSED | Total investment, not per-token economics |
| 2025 AI revenue / op loss | $3.201B / −$6.355B | DISCLOSED | |
| 2025 / Q1 2026 AI capex | $12.727B / $7.723B | DISCLOSED | |
| Jukan C1 rental equivalent | $2.60/GPU-hr | COMMUNITY ESTIMATE | Modeled lease, not owned cost |
| Reported training MFU | ~11% | CREDIBLY REPORTED | Training only |
| Anthropic capacity price | ~$5.27/GPU-hr | derived from DISCLOSED | Opportunity cost of Grok serving |
| Google future capacity price | ~$11.45/GPU-hr | derived from DISCLOSED | Future, not current |
Central: Grok 4.5 marginal serving gross margin at API list ≈ 67% — SPECULATION. Subjective 80% CI: ~10% to 85% — SPECULATION.
8-GPU GB300 replica; 80 tok/s user stream (DISCLOSED); 25 concurrent streams at saturation (SPECULATION) → 2,000 aggregate tok/s/replica → 0.004 GPU-sec/output token; $2.40/GPU-hr economic TCO → $2.67/M output cost; input $0.40/M uncached, $0.10/M cached.
| Unit | List revenue | Central cost | Central margin |
|---|---|---|---|
| 1M output | $6.00 | $2.67 | 55.6% |
| 1M uncached input | $2.00 | $0.40 | 80.0% |
| 1M cached input | $0.50 | $0.10 | 80.0% |
| 3M uncached in + 1M out | $12.00 | $3.87 | 67.8% |
| Same with 50% input cached | $9.75 | $3.42 | 65.0% |
At strict cash-marginal $0.60/hr (hardware sunk): ~$0.67/M output → ~92% margin on the 3:1 workload. At the Anthropic-contract $5.27/hr opportunity value: ~$5.86/M output → 2% output-only margin, ~29% blended — serving output-heavy traffic would barely beat selling the capacity wholesale.
Clean interpretation: Zephyr's "priced closer to cost" implication is directionally supportable under full-cycle economic TCO, but not under strict cash marginal cost. xAI is not obviously pricing at 90–95% full-cycle margin; it may nevertheless earn roughly that on the next token whenever spare, already-paid-for capacity is available.
| GPU valuation | Break-even aggregate throughput | Equivalent 80-TPS streams |
|---|---|---|
| $0.60/hr cash | 204 tok/s | 2.6 |
| $2.40/hr TCO | 889 tok/s | 11.1 |
| $5.27/hr Anthropic value | 1,952 tok/s | 24.4 |
Central case (25 streams / 2,000 tok/s) is comfortably profitable against owned TCO but almost exactly break-even against the Anthropic capacity alternative.
Monte-Carlo-style run: 10th/50th/90th percentile blended margins ≈ 10% / 62% / 84%.