← Frontier Inference Margins · all research reports

Unedited research artifact (2026-07-09) — published as reference because its inline citations are useful; conclusions are synthesized (and where needed corrected) in the main report.

I'll search X extensively for TeorTaxes, Jukan, Zephyr, Musk's Opus claim, and related serving-cost threads.# X/Twitter research report: frontier inference margins & Anthropic cost-to-serve

Search date: 2026-07-09. Every claim below is tied to a direct post URL, handle, date, and a verbatim quote. Numbers are not paraphrased.


1. @teortaxesTex — high inference margins + DeepSeek 545% / H800 economics

Strongest posts with numbers (5)

# URL Author Date (UTC) Verbatim quote Numbers claimed
1 x.com/teortaxesTex/status/2071314837771014298 @teortaxesTex 2026-06-28 "…no, they can't have 90%+ margins? Right? Right?" THEY CAN. 90%+ margins (frontier/west labs context; reply to skepticism that western labs have “insane margins”)
2 x.com/teortaxesTex/status/2071315379004051513 @teortaxesTex 2026-06-28 math for DeepSeek Serving Opus is at most $4/1Mt. Anthropic is not retarded They are fleecing you even on subscriptions, except maybe at 100% utilization Serving Opus ≤ $4/1Mt; subscription “fleecing” except ~100% utilization
3 x.com/teortaxesTex/status/2070786814097440805 @teortaxesTex 2026-06-27 No, they'll just increase the batch size, have the same speed, and drive margins from 90% to 95%. You're welcome Western providers: 90% → 95% margins if they keep speed and raise batch size (reply about DeepSeek-driven throughput fixes)
4 x.com/teortaxesTex/status/2070833810145415518 @teortaxesTex 2026-06-27 I see 70-90 tps right now, so let's say they're doing 1.4K t/s/gpu. That's 5M tokens/hour, or ≈$4,4 at $0.87/Mt. H100 *spot prices* are now $2.4/hr. DS has at least 50% margin. 70–90 tps; 1.4K t/s/gpu; 5M tokens/hour; ≈$4.4 at $0.87/Mt; H100 spot $2.4/hr; DS ≥50% margin
5 x.com/teortaxesTex/status/1906521148726100430 @teortaxesTex 2025-03-31 if we exclude R&D and look at inference alone, Anthropic and OpenAI are making like 80% margins. Inference-only: Anthropic & OpenAI ~80% margins

Related TeorTaxes posts (supporting / DeepSeek 545% reactions)

URL Author Date Verbatim quote Numbers
x.com/teortaxesTex/status/1895876441117769922 @teortaxesTex 2025-03-01 Thanks. What is OpenAI's profit margin? (quoting Adam: “It’s an 84% profit margin. Impressive, but not 545%.”) Contests reframing of DeepSeek’s 545% as 84%; asks OpenAI’s margin
x.com/teortaxesTex/status/1895876331495477615 @teortaxesTex 2025-03-01 nice lesson in arithmetic Adam, but could you tell us your profit margin? Same thread
x.com/teortaxesTex/status/1895889315970044153 @teortaxesTex 2025-03-01 very embarrassing error for o3-mini-high, it's larping hard here (quoting o3-mini on $562,027/day revenue, $87,072/day cost, 545% vs ~84%) Critiques 545% → 84% “correction” math
x.com/teortaxesTex/status/1896141536884396528 @teortaxesTex 2025-03-02 What they technically mean is that they have a markup of 545% Clarifies 545% as markup, not profit margin
x.com/teortaxesTex/status/2070832301005537627 @teortaxesTex 2026-06-27 DeepSeek has disclosed their inference economics – for the first time since Open Source Week. Then, they were doing 14.8K generation on 8xH800 node (1,85K/GPU) at 20-22 tps. Whatever these GPUs are now, V4-Pro is *at least 3x cheaper to serve*. 2K/GPU needs >60 tps. Open Source Week: 14.8K gen / 8×H800 (1.85K/GPU) at 20–22 tps; V4-Pro ≥3× cheaper to serve; 2K/GPU needs >60 tps
x.com/teortaxesTex/status/1901766584974155784 @teortaxesTex 2025-03-17 DeepSeek has provided their inference figures. > Each H800 node delivers an average throughput of ~73.7k tokens/s input … or ~14.8k tokens/s output … That's at 20ish t/s though Quotes DS: ~73.7k in / ~14.8k out per H800 node; ~20 t/s
x.com/teortaxesTex/status/2068833223228924229 @teortaxesTex 2026-06-21 They do have 90% margins 90% margins (short reply; limited parent context in search)

Primary DeepSeek disclosure (source TeorTaxes was reacting to)

URL Author Date Verbatim quote Numbers
x.com/deepseek_ai/status/1895688300574462431 @deepseek_ai 2025-03-01 Day 6 of #OpenSourceWeek… 73.7k/14.8k input/output tokens per second per H800 node… Cost profit margin 545% 73.7k/14.8k t/s per H800 node; cost-profit margin 545%
x.com/Yuchenj_UW/status/1895690693617799449 @Yuchenj_UW 2025-03-01 DeepSeek is able to get 73.7k tokens/s input and 14.8k tokens/s output throughput per H800 node! Their profit margin is 545%… Same figures, widely amplified
x.com/reach_vb/status/1895716781139853364 @reach_vb 2025-03-01 Cost: $87,072 daily… Theoretical Revenue: $562,027 daily (545% margin) $87,072/day cost; $562,027/day theoretical revenue; 545%
x.com/Beth_Kindig/status/1895716781139853364 is wrong ID — correct: x.com/Beth_Kindig/status/1895871653793427955 @Beth_Kindig 2025-03-01 DeepSeek revealed that its AI models have up to a 545% cost-profit ratio, with daily inference costs of ~$87,000… daily revenue of ~$562,000. 545% cost-profit; ~$87k cost; ~$562k revenue

Note on TeorTaxes 90–95% claim: He repeatedly asserts 90%+ / 90→95% / ~80% inference-only for Anthropic/OpenAI/western serving, with the sharpest numeric Opus cost claim being “Serving Opus is at most $4/1Mt” (2026-06-28). I did not find a single TeorTaxes post that says exactly “Anthropic marginal gross margin is 92%” as a firm audited figure—his language is “90%+”, “90% to 95%”, “like 80%”, and “THEY CAN.”


2. Jukan & Zephyr — handles + similar margin/serving claims

Handle corrections

Your guess Status Correct account
@Jukanlosreve Wrong for this person. @AveryAdrial uses display name “Jukanlosreve” but has ~54 followers and is not the AI/chips account. @jukan05 — “Jukan @ ICML” (high-follower AI/semiconductor poster; bio: Tech otakus save the world | Not Investment Advice)
@zephyr_z9 Correct @zephyr_z9 — “Zephyr”, ~161k followers; bio: AI & Chips | Not Investment Advice | DYOD

Zephyr (@zephyr_z9) — direct margin / serving economics posts

URL Author Date Verbatim quote Numbers
x.com/zephyr_z9/status/2074917201589727588 @zephyr_z9 2026-07-08 At least xAI isn't juicing up the gross margins to 90%-95% and scamming consumers Although the cached token could be cheaper Frontier/Anthropic-style gross margins 90%–95% (implicit contrast with xAI Grok 4.5 pricing)
x.com/zephyr_z9/status/2069832997218722140 @zephyr_z9 2026-06-24 70% GM and 15%-20% FCF margin Damn Anthropic is printing cash now Anthropic: 70% GM, 15%–20% FCF margin
x.com/zephyr_z9/status/2070726279540989971 @zephyr_z9 2026-06-27 I think they will do over $110B in revenue next year at 75%-80% GM >$110B revenue next year; 75%–80% GM (context in thread appears Anthropic/frontier — confirm parent if using)
x.com/zephyr_z9/status/2070727375386718259 @zephyr_z9 2026-06-27 $50B-$60B this year over $110B next year 75%-80% GM $50–60B this year; >$110B next; 75%–80% GM
x.com/zephyr_z9/status/2048029866633895944 @zephyr_z9 2026-04-25 I'm pretty sure V4 Pro is cheaper to serve than They went from 90% margin to 50% margin DeepSeek V4-Pro promo: 90% → 50% margin (quoting TeorTaxes breakeven pricing thread)
x.com/zephyr_z9/status/2071528823422959737 @zephyr_z9 2026-06-29 The production cost of those 70M tokens are less than $15 (quoting ~$268–$286 list value for 65–70M tokens) 70M tokens production cost <$15 vs ~$268–$286 list
x.com/zephyr_z9/status/2074755502878470238 @zephyr_z9 2026-07-08 Semianalysis is forecasting that Anthropic is processing 5Q tokens per month in June, which will grow to 13Q tokens by December Anthropic volume: 5Q tokens/month (June)13Q (December)
x.com/zephyr_z9/status/2075107995508408728 @zephyr_z9 2026-07-09 Nah Opus/Fable have the highest active parameters OAI is more sparse (still around 100B active range) ~100B active for OAI; Opus/Fable highest active params
x.com/zephyr_z9/status/2070267916013187449 @zephyr_z9 2026-06-25 That's around 25 trillion Tokens… assuming, 12:1 input: output ratio and 90% caching on Opus So it's nearly around $72M and around $160M if no caching is assumed Distillation abuse estimate: 25T tokens; 12:1 I/O; 90% caching; ~$72M / ~$160M without cache

Jukan (@jukan05) — hardware/serving economics (not 90–95% Anthropic API margins)

I did not find posts from @jukan05 claiming Anthropic or frontier-lab token inference gross margins of 90–95%. His widely cited posts are about GPU/memory/cluster economics and Anthropic compute leasing, not unit-token COGS.

URL Author Date Verbatim quote Numbers
x.com/jukan05/status/2052957921563316619 @jukan05 2026-05-09 xAI deployed more than 220,000 NVIDIA GPUs… rented out at roughly $2.60 per GPU-hour… minting $5–6 billion in annual revenue… Anthropic… converts roughly $5B of spend into what it expects to be $15B of ARR (long thread; also earlier note: “xAI can recognize $6B of annual revenue… Anthropic… $5B of spend into… $15B of ARR”) 220k+ GPUs; ~$2.60/GPU-hr; $5–6B annual lease revenue; Anthropic ~$5B spend → ~$15B ARR
x.com/jukan05/status/2071139712082022855 @jukan05 2026-06-28 Inference is now directly tied to money… GPUs are often underutilized in inference… For inference, adding more memory is far more valuable… Ultimately, the ROI of inference depends less on GPUs and more on memory… Inference is memory. Inference ROI = memory-bound, not GPU-count
x.com/jukan05/status/2073032040451366952 @jukan05 2026-07-03 Bernstein estimates that SK hynix’s DRAM gross margin in Q2 will reach 90.9%… 90.9% DRAM GM (not Anthropic)
x.com/jukan05/status/2071745275325190339 @jukan05 2026-06-29 commodity DRAM operating margins already at 80% and gross margins above 90% DRAM OM 80%, GM >90%

3. Elon Musk — Claude Opus parameter “leak”

Important correction to the prompt framing

The viral claim is not that Opus is much smaller than expected. Elon said current Grok is small, and Opus is ~10× Grok, which the timeline widely read as Opus ≈ 5T total parameters (large). The “surprise” was Grok being strong for 0.5T, not Opus being tiny.

Primary Elon post

URL Author Date Verbatim quote Numbers / implication
x.com/elonmusk/status/2042123561666855235 @elonmusk 2026-04-09 0.5T total. Current Grok is half the size of Sonnet and 1/10th the size of Opus. Very strong model for its size. Grok 0.5T total; Grok = ½ Sonnet, 1/10 Opus → community deduction: Sonnet ≈ 1T, Opus ≈ 5T

Context: Reply in a thread about Grok 4.2 / training fleet sizes (parent discussion referenced Colossus models in training including multi-T variants). Not framed as xAI litigation discovery; it is a public tweet sizing Grok vs Anthropic models. No courtroom document link was attached in the post itself.

Notable reaction posts (size + cost implications)

URL Author Date Verbatim quote Numbers
x.com/lifetimization/status/2042285703162397167 @lifetimization 2026-04-09 lol Elon musk leaked Sonnet’s and Opus’s sizes: Sonnet: 1T Opus: 5T Sonnet 1T, Opus 5T
x.com/mark_k/status/2042304980456030666 @mark_k 2026-04-09 Elon Musk has casually leaked the model sizes… Grok 4.20, which is 0.5T parameters. Now we can deduce: - Claude Sonnet = 1T - Claude Opus = 5T Same deduction
x.com/Mayhem4Markets/status/2042306488769683712 @Mayhem4Markets 2026-04-09 how do you know the size of Sonnet and Opus? … Sonnet is a 1T parameter model and Opus is a 5T parameter model. Questions source; same 1T / 5T read
x.com/teortaxesTex/status/2068980091724419280 @teortaxesTex 2026-06-22 Seeing as Fable was served at like 90 t/s… it seems Fable had shockingly FEW active parameters for what it was. (quoting his earlier: If Mythos is above, say, 800B active…) ~90 t/s serving; Mythos >800B active as resignation threshold; few active params on Fable (active, not total)
x.com/teortaxesTex/status/2071325290819641620 @teortaxesTex 2026-06-28 Opus has at most 2-4x active params do some actual quantitative math for once, germoid Opus active params ≤2–4× (relative to interlocutor’s baseline—not absolute B)

What I did not find


4. Other widely-cited 2025–2026 X threads on Anthropic cost-to-serve / margins

URL Author Date Verbatim quote Numbers claimed
x.com/PodcastAlphaX/status/2072119494563262697 @PodcastAlphaX 2026-07-01 Dylan Patel @dylan522p of SemiAnalysis: Anthropic's margin on an Opus 4.8 API token is north of 80%. … At 80%-plus, even doubling compute costs leaves Anthropic above 50% gross margin. Opus 4.8 API token margin >80%; after 2× compute still >50% GM
x.com/SemiAnalysis_/status/2037575752636301499 @SemiAnalysis_ 2026-03-27 Anthropic's gross margins were -94% in 2024. MiniMax was -25%. … later: we think Inference Provider Gross Margins should blend to ~60%.Labs that understand the interactivity lever operate at 60%+ margins. Company GM −94% (2024) vs blended provider target ~60% / 60%+ (accounting GM ≠ unit inference margin)
x.com/IvanaSpear/status/2075227080657129631 @IvanaSpear 2026-07-09 Inference costs fell from ~$10.86 per million tokens in early 2024 to ~$0.25 today (roughly 40x) while revenue per token fell only ~9x. … Quarterly gross profit swings from -$55M to an estimated $453M, with 3Q26 profit crossing $1B. … Claude has reportedly been gross-margin positive since day one. … training costs fall from 400%+ of revenue to an estimated ~36% by year-end. $10.86 → $0.25 /M tokens (~40×); rev/token ~9× down; GP −$55M → ~$453M; 3Q26 profit >$1B; training 400%+ → ~36% of revenue
x.com/kimmonismus/status/2014673235594641838 @kimmonismus 2026-01-23 Anthropic slashed its 2025 gross margin outlook to 40% as inference costs came in 23% higher than expected, even while revenue is projected to hit $4.5 billion 2025 GM outlook 40%; inference costs +23%; rev $4.5B
x.com/fleetingbits/status/2073528885149679622 @fleetingbits 2026-07-04 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 Frontier inference margins ~40–50% (counter to 90% narrative)
x.com/rosyprosperity/status/2073086578616836598 @rosyprosperity 2026-07-03 Anthropic is almost certainly running gross margins of over 90% on the frontier model inference. >90% GM on frontier model inference
x.com/SemiAnalysis_/status/2060043702450446552 @SemiAnalysis_ 2026-05-28 The most popular AI subscription will run you about $20/month… for a company like Anthropic how much does it cost the company to be servicing the user? … the same $20 subscription can range from insanely profitable to barely breaking even. $20/mo sub; cost-to-serve depends on workload (no single $/M in this post text)
x.com/aakashgupta/status/2020272837907669179 @aakashgupta 2026-02-07 standard Opus 4.6 runs $5/$25 per million tokens. Fast mode runs $30/$150. A 6x premium. … capture full margin instead of sharing 30-40% with AWS and Google. List prices $5/$25 vs Fast $30/$150 (); cloud take 30–40%
x.com/deepseek_ai/status/1895688300574462431 @deepseek_ai 2025-03-01 (see §1) Open Source Week inference system Benchmark open disclosure used as yardstick for all labs

Narrative tension worth flagging (not resolving)

Camp Representative claim Example
High unit / API margins 80–95% on frontier tokens / “juicing GM to 90–95%” TeorTaxes, Zephyr, Dylan Patel clip, @rosyprosperity
Mid reported company GM ~40–70% company gross margin / 40–50% “frontier lab inference margins” @kimmonismus (40% outlook), @fleetingbits (40–50%), Zephyr’s 70% GM Anthropic cash-print post
Historical company GM −94% Anthropic GM in 2024 SemiAnalysis Mar 2026 thread

These are different objects (unit token margin vs GAAP gross margin vs theoretical DeepSeek cost-profit ratio). Search results do not show consensus on one number.


Gaps (explicit non-findings)

  1. @Jukanlosreve is not the AI analyst account; use @jukan05.
  2. @jukan05 posts on Anthropic cluster lease / memory / DRAM economics were found; not a clear Jukan thread asserting 90–95% Anthropic inference GM.
  3. Elon did not claim Opus is small. He claimed Grok is 0.5T and 1/10 of Opus → public deduction Opus ~5T. No litigation-document “leak” of a small Opus param count turned up in this search.
  4. No primary post found with a fully sourced, end-to-end tokens/sec/GPU → $/M tokens COGS → margin table for Anthropic Opus with hardware SKUs and utilization assumptions comparable to DeepSeek’s Open Source Week disclosure. Closest numeric serving claims: TeorTaxes Opus ≤ $4/1Mt, SemiAnalysis/Dylan >80% Opus 4.8 API margin, IvanaSPEAR $10.86 → $0.25 /M cost collapse.

Quick citation index (strongest single posts per theme)