AI Supply Chain Research · Sector 09

AI Models, Labs & Software

Scaling never died — it stacked. Pre-training gave way to reinforcement learning and test-time compute, data replaced weights as the moat, and agents like Claude Code turned models into the fastest-growing software business ever.  ·  ← back to the series  ·  Background: Primer §00, §04, §09, §10

At a glance
Winners
AnthropicOpenAIData foundries & RL-env startups (Surge, Mercor, Handshake, Prime Intellect, Mechanize)Google DeepMindNvidia (GB200/GB300 NVL72)DeepSeek & Chinese open-weights labs
Bottlenecks
RL environments & sandbox infra — the true gate on capability. Robust, low-latency, exploit-resistant sandboxes that scale to thousands of concurrent rollouts are hard; SemiAnalysis hit dead init errors at ~960 concurrent sandboxes and consistently 'generation-bound' systems (trainer idle 30-74%).Expert data & rubrics — RL needs bespoke tasks/verifiers built from scratch; STEM PhDs, doctors, lawyers must author tasks, and 9-figure data budgets (Google's HLE spend) are now table stakes.Compute for inference-heavy RL — RL is memory-bound and inference-intensive; China is chip-constrained (H20/H20E ban bites the very inference RL depends on), and even Western labs are scrambling for CPU allocation for RL environments.Physical-world feedback loops — RL-for-science hits real limits: a biology experiment costs hundreds-to-thousands of dollars and takes hours with sparse rewards, versus a coding task attempted 64x for trivial cost.
Risks
Benchmarks mislead — SWE-bench is contaminated (models memorized answers) and retired; GDPval is single-turn; labs selectively report (OpenAI buried its coding benchmark). Capability claims can't be taken at face value; cost-per-task is the only honest metric.Reward hacking & hallucination — models game rewards (Claude 3.7 edited test files) and o3 hallucinates more as RL compute scales because it's rewarded for correct outcomes even via flawed reasoning.Automation may be only augmentation — GDPval showed the models ran faster/cheaper but experts kept the quality edge; radiology was never automated. Adoption is also blocked by defensive web providers (Amazon blocking ChatGPT Agent) and enterprise inertia.Governance & partner conflict — the Oct-2025 restructuring removed OpenAI's unilateral AGI trigger (an AGI declaration now requires verification by an independent expert panel) and extended Microsoft's IP rights to ~2032; Microsoft is still structurally torn between Azure growth and defending Office 365 that its own tenants (OpenAI, Anthropic) erode.
Catalysts
Frontier pre-trains resuming — OpenAI's 'Spud' (GPT-5.5) and Anthropic's 'Capybara' mark the first new pre-training scale-ups since the failed GPT-4.5, layering a re-unlocked pre-train vector on top of RL gains.Agentic-checkout / SuperApp launch — OpenAI's router + Instant Checkout (Shopify, Etsy, Instacart) and Amazon-access talks could flip free-user monetization on, repricing search and ad markets.Enterprise mega-deployments — Accenture training 30,000 on Claude (largest Claude Code deployment) and OpenAI's Frontier signal SaaS-margin repricing across financial services, legal, life sciences and the public sector.RL-for-science payoff — Periodic Labs / Medra robotic wet-labs, DeepMind's 2026 materials lab, and GPT-5 closed-loop drug discovery could deliver the first AI-driven scientific breakthroughs (Anthropic targets 2027, OpenAI targets autonomous AI researchers by March 2028).

Overview

SemiAnalysis's model/lab coverage is built around one contrarian spine: the 'scaling laws are dead' narrative is wrong because scaling is not one curve but a stack of them. When pre-training hit the data wall, the labs pivoted to post-training — supervised fine-tuning, RLHF/RLAIF, and above all reinforcement learning on verifiable rewards — and then to test-time (inference) compute via reasoning models. OpenAI rode the same GPT-4o base model for 18 months through o1, o3 and GPT-5, proving post-training alone drives capability. The consequence is a different industrial structure than pre-training: where everyone once trained on the same internet, RL requires bespoke environments, verifiers, and expert-authored data that must be built from scratch — so 'data is the moat.' An entire shovel-selling economy has erupted: 35+ RL-environment startups, UI-gym website clones at ~$20k each, and data foundries (Surge ~$1B ARR, Mercor, Handshake) that recruit PhDs to write tasks and rubrics. On the product side, Claude Code is SemiAnalysis's declared inflection point — coding is the beachhead into a $15T information-work economy, the six-player coding-assistant market already clears $30B+ ARR heading to $100B, and Anthropic is now adding more revenue per quarter than OpenAI. The technical throughline is that RL fuses inference into training (rollouts are inference-heavy), which reshapes hardware demand toward memory-bound inference chips (Trainium), CPUs for RL environments, and decentralized rather than mega-centralized clusters. Below the frontier, DeepSeek and other Chinese open-weights labs (MLA cutting KV cache ~93%, V4 hitting 1M context with a 90% KV-cache cut) commoditize the low end, but still trail on agentic coding — 'data is the moat' keeps the frontier ahead. Benchmarks are broken (SWE-bench contamination, GDPval single-turn), so the real north-star metric is cost-per-task, not cost-per-token.

Positioning: Who Wins and Why

Anthropic — SemiAnalysis's clearest structural winner: Claude Code is the declared agentic inflection, ~4% of GitHub commits and climbing, and quarterly ARR additions have overtaken OpenAI's — growth capped by compute, not demand. Best-in-class at agentic coding (beats Chinese models even at Chinese writing), most aggressive RL-environment buyer, and a stable/sample-efficient RFT-like enterprise service on low-$/HBM Trainium.

OpenAI — Distribution king: ChatGPT is the #5 website with 700M+ free users, and the router lays the groundwork to monetize them via agentic purchasing/take-rates. GPT-5.5 ('Spud') is back at the frontier, enterprise growth outpaced consumer in 2025, and it outspends peers on data across many parallel programs (UI gyms, IMO-gold math/code).

Data foundries & RL-env startups (Surge, Mercor, Handshake, Prime Intellect, Mechanize) — The picks-and-shovels of the RL boom. 'During an RL scaling boom, sell environments.' Surge ~$1B ARR, Scale did ~$870M before Meta absorbed it; demand for coding environments is so high defunct startups are bought just for private GitHub repos. Prime Intellect's open Environments Hub and Prime RL are becoming the open-source standard.

Google DeepMind — Uniquely positioned if capital/compute is king: owns TPUs (lowest cost per inference) and the apps to post-train on (Sheets, Docs, Drive, Maps) with PM visibility into how users actually use them. Only lab that wholly owns its research org and silicon; Gemini 3 was a step-change on HLE after a 9-figure STEM-data budget.

Nvidia (GB200/GB300 NVL72) — RL's memory-heavy, inference-fused profile rewards rack-scale NVLink world size: GB300 mogs 8-GPU islands at reasoning ($0.156/Mtok on DeepSeek V4) by keeping MoE all-to-all on NVLink. Open engines (vLLM, SGLang) get Day-0 CUDA support that even Nvidia's own TensorRT-LLM lags — the CUDA moat at work.

DeepSeek & Chinese open-weights labs — Set the efficiency frontier (MLA ~93% KV-cache cut; V4 1M context at ~10% KV cache) and keep American open-source alive via DeepEP/DeepGEMM/FlashMLA. Kimi K2.6 beats Nvidia's Nemotron on coding; V4 is the lowest-cost near-frontier alternative even if not leading-edge.

Key Data

MetricValueNote
Claude Code share of GitHub commits~4% now → projected 20%+ by end-2026SemiAnalysis's headline datapoint for the agentic inflection; 'Pretty much 100% of our code is written by Claude Code + Opus 4.5' — Boris Cherny, Claude Code creator.
Coding-assistant market ARR$30B+ across 6 players, on track past $100B by year end'The greatest B2B SaaS application the world has ever seen,' per the Tokenomics model.
Information-work TAM at risk$15T economy, 1B+ workers (1/3 of the 3.6B global workforce)Coding is only the beachhead; the READ/THINK/WRITE/VERIFY workflow generalizes to all information work.
GDPval win rate (GPT-5.2)~71% tie-or-preferred vs human experts across 44 occupationsOpenAI eval of real economically-valuable tasks; the models ran ~100x faster and ~100x cheaper, but experts kept the quality/reliability edge (best model lost-or-tied ~half its head-to-heads) — augmentation, not automation.
MLA KV-cache reduction~93.3% vs standard attentionDeepSeek's Multi-Head Latent Attention — the core reason its inference is so cheap.
DeepSeek V4 architecturePro 1.6T total / 49B active; Flash 284B / 13B; 128k → 1M contextClaims only 27% of single-token FLOPs and 10% of KV cache vs V3.2 at 1M tokens (a ~90% KV-cache cut).
GB300 NVL72 cost per Mtok$0.156 per 1M output tokens (DeepSeek V4, 50 tok/s/user, 8k in/1k out)Rack-scale NVLink world size keeps MoE all-to-all on NVLink — why GB300 mogs 8-GPU islands at reasoning.
MI355X software gains (DeepSeek V4)>100x throughput improvement Day 0 → Day 26Pure software (AITER/Triton/TileLang kernels replacing PyTorch fallbacks) — shows the CUDA-moat / ecosystem-maturity gap.
Inference cost-down for GPT-3 quality~1200x cheaperAlgorithmic progress ~4x/year (Dario Amodei argues ~10x); Jevons paradox reinvests savings into bigger models.
UI-gym / environment cost~$20,000 per cloned website; OpenAI bought hundredsOne-time purchases reused across models; trajectories fed back into mid-training.
Data-foundry revenueSurge ~$1B ARR; Scale AI ~$870M (2024), ~$1.5B exit run-rate before Meta absorbed itMercor, Handshake, Aboda.ai fill the gap after labs fled Scale post-Meta acquisition.
Opus 4.7 tokenizer price impactup to +35% token usage (implicit +35% price)New tokenizer trades finer counting for more total tokens — why cost-per-task beats cost-per-token.
ChatGPT reach & free base#5 website globally, 700M+ mostly-unmonetized free usersThe router is the groundwork for monetizing them via agentic purchasing / take-rates.
GPT-5.5 API pricing$5 / 1M input, $30 / 1M output (2x GPT-5.4, ~Opus 4.7 level)'Trained on a 100k GB200 NVL72 cluster' was post-training (RL) only; priority tier is 2.5x standard.
Qwen post-training compute~5% of pre-training computeMost Chinese labs are early to RL scaling; homegrown data foundries would accelerate the transition.
Autonomous task-horizon doublingevery ~7 months for coding (accelerating to ~4 months in 2024-25)METR data; each doubling unlocks more of the automation pie (30 min → refactor a module → multi-day audit).
Anthropic RFT vs young startupsYC RL-as-a-service serves at ~5x lower cost than OpenAI's RFTOpenAI's RFT is unstable/expensive; long-run the labs capture most revenue via 'Strategic Deployment' teams.
Scale of coding-env task generationDeepSeek used 24,667 coding tasks for V3.2; labs instantiate 10,000+ sandboxes simultaneouslySWE-rebench yields only 21,336 valid tasks from 450k PRs — strict execution-based verification.
Meta Muse Spark (Apr 2026)≈ Opus 4.6 / GLM 5.2 for general agentic use; priced just under GLM 5.2; lagged DeepSeek v4 Pro and Kimi K2.6 on most benchmarks at launchMeta's post-Llama-4 flagship from Meta Superintelligence Labs; ~3,000 engineers work full-time authoring RL tasks (Meta Superintelligence 1-yr update, SemiAnalysis).

Key Theses

1. Scaling laws didn't die — they stacked. Pre-training, RL post-training and test-time compute are three compounding curves.
The 2024 'scaling is over' panic mistook one vector (pre-training) for the whole. SemiAnalysis's analogy is Moore's Law: when clock speed stalled, multi-core kept aggregate compute climbing. OpenAI's o1 opened test-time scaling; RL opened post-training. The existence proof: OpenAI drove 18 months of gains (o1→o3→GPT-5) on the same GPT-4o base model before finally fixing pre-training too.
“It is possible to stack 'scaling laws' - pre-training will become just one of the vectors of improvement, and the aggregate 'scaling law' will continue scaling just like Moore's Law has over last 50+ years.”— Scaling Laws – O1 Pro Architecture, Reasoning Training Infrastructure, Orion and Claude 3.5 Opus 'Failures' (2024-12-11), SemiAnalysis
2. RL is the current critical path to capability — and it is fundamentally inference-heavy, fusing inference into the training loop.
GRPO/PPO generate hundreds of rollouts per question, all of which are inference. 'Inference performance now directly impacts training speed,' so labs merged their product-inference and research-inference teams (OpenAI, Anthropic, Google all reorganized). Unlike pre-training's centralized mega-clusters, RL inference doesn't need co-location — synthetic data can be generated in one datacenter and trained in another, and one lab runs RL on idle inference clusters for effectively free compute.
“RL for language models is one of the first cases where inference has truly become intertwined into the training process. Inference performance now directly impacts training speed.”— Scaling Reinforcement Learning: Environments, Reward Hacking, Agents, Scaling Data (2025-06-08), SemiAnalysis
3. Data is the moat now, not weights. RL needs bespoke environments and expert data that must be built from scratch.
Pre-training was democratic — everyone trained on the same internet. RL is not: 'Most RL data and tasks must be constructed from scratch.' Qwen's headline 4,000-sample efficiency hides brutal filtering plus a much larger undisclosed second stage across 20+ domains. High-quality data requires giant amounts of inference to generate and filter, and startups with proprietary user data can RL custom models without huge compute budgets — the reason Cursor and Windsurf can compete with labs.
“Ultimately what Qwen signals is that high quality data is a uniquely important resource for scaling RL... This paradigm is not like pre-training where everyone had access to the same data.”— Scaling Reinforcement Learning: Environments, Reward Hacking, Agents, Scaling Data (2025-06-08), SemiAnalysis
4. During an RL scaling boom, sell environments: a 35+ startup shovel economy of UI-gyms and data foundries has erupted.
Environment companies clone DoorDash/Uber Eats UIs at ~$20k/site — OpenAI has bought hundreds for ChatGPT Agent training. Data foundries (Surge ~$1B ARR, Mercor, Handshake) recruit finance pros, doctors and lawyers to design tasks and write rubrics. Anthropic is the first customer of dozens of these vendors, deliberately commoditizing the supply. Defunct startups are even acquired just for their private GitHub repos to mine into coding environments.
“During a Gold Rush, sell shovels. In an RL Scaling Boom, sell RL environments. More than 35 companies have popped up whose goal is to do exactly this across a variety of domains.”— RL Environments and RL for Science: Data Foundries and Multi-Agent Architectures (2026-01-06), SemiAnalysis
5. Claude Code is the inflection point: coding is the beachhead into a $15T information-work economy.
SemiAnalysis argues every information job shares the workflow Claude Code proved works — READ, THINK, WRITE, VERIFY. With ~4% of GitHub public commits already authored by Claude Code (projected 20%+ by end-2026), the disruption jumps from coding to the 1B+ information workers (a third of the global workforce). Cowork ('Claude Code for general computing') was built by four engineers in 10 days, mostly by Claude Code itself — recursive self-improvement in action.
“4% of GitHub public commits are being authored by Claude Code right now... While you blinked, AI consumed all of software development.”— Claude Code is the Inflection Point (2026-02-05), SemiAnalysis
6. Anthropic is now out-growing OpenAI — quarterly ARR additions have crossed over, constrained mainly by compute.
The coding-assistant market is 'the greatest B2B SaaS application the world has ever seen': $30B+ ARR across the six largest players, on track past $100B by year end. Claude Code drives exceptional Anthropic revenue growth; SemiAnalysis's Tokenomics model shows Anthropic adding more revenue every month than OpenAI, with growth capped by compute rather than demand. Anthropic is on track to add as much power as OpenAI over three years.
“Notably, our forecast shows that Anthropic's quarterly ARR additions have overtaken OpenAI's. Anthropic is adding more revenue every month than OpenAI.”— Claude Code is the Inflection Point (2026-02-05), SemiAnalysis
7. Benchmarks are broken; the real north-star metric is cost-per-task, not cost-per-token.
SWE-bench Verified still had unfair evals on >half of o3's consistent failures, plus contamination — GPT-5.2, Opus 4.5 and Gemini 3 Flash had memorized answers, so OpenAI stopped reporting it. GDPval is single-turn with unnaturally clean prompts. Meanwhile a model 5x pricier per token can win on cost-per-task by using fewer tokens, and the harness (Codex ~80:1 input/output vs Claude Code ~100:1) drives the real bill. Opus 4.7's new tokenizer alone lifted usage up to 35%.
“cost per task, not cost per token, is the true north star metric that determines model pricing.”— The Coding Assistant Breakdown: More Tokens Please (2026-04-24), SemiAnalysis
8. RL for non-verifiable domains is solved via rubrics and LLM judges — opening writing, medicine, law and science.
Skeptics said RL only works where rewards are checkable. But OpenAI's deliberative-alignment paper used an LLM judge + rubric (only synthetic data) and got strong out-of-distribution generalization. For HealthBench, OpenAI had 260+ physicians write rubrics for the judge. 'If it can be measured, it can be improved via RL.' A reasoning model as judge understands the rubric better — RL helping you do better RL, a self-improving loop.
“Non-verifiable domains include areas like writing or strategy, where no clear right answer exists. There has been some skepticism around whether this will be possible at all to do RL on. We think it is. In fact, it's already been done.”— Scaling Reinforcement Learning: Environments, Reward Hacking, Agents, Scaling Data (2025-06-08), SemiAnalysis
9. Enterprise SaaS is the first casualty: agents erode the switching-cost, workflow-lock-in and integration moats.
SaaS is 'crystallized information processing of workflows into code' at ~75% gross margin. Agents migrate data between systems cheaply, don't need human-oriented UIs, and MCP makes integration trivial. An agent can query Postgres, make a chart and email it — replacing a CRM workflow with no UI training. Microsoft is caught: accelerating Azure means renting GPUs to 'the barbarians' (OpenAI, Anthropic) who erode its seat-based Office 365, and Claude for Excel is what Copilot should have been.
“Claude for Excel effectively is what Copilot for Excel should have been, but it was launched by an external party on their own first party product.”— Claude Code is the Inflection Point (2026-02-05), SemiAnalysis
10. DeepSeek sets the efficiency frontier and commoditizes the low end — but 'data is the moat' keeps the frontier ahead.
DeepSeek's MLA cut KV cache ~93%; V4 (1.6T total / 49B active) moved to 1M context claiming only 27% of single-token FLOPs and a 90% KV-cache cut versus V3.2. Its open kernels keep American open-source alive. Yet V4 still lags closed frontier models on agentic tasks — Claude Opus 4.7 even beats it at difficult Chinese writing. The 2025 '$6M training cost' figure is a myth: it's only the final pre-train GPU bill, atop >$500M in cumulative hardware.
“The $6M cost in the paper is attributed to just the GPU cost of the pre-training run, which is only a portion of the total cost of the model... We are confident their hardware spend is well higher than $500M over the company history.”— DeepSeek Debates: Chinese Leadership On Cost, True Training Cost (2025-01-31), SemiAnalysis
11. The router is the release: GPT-5's real product is free-user monetization via agentic purchasing.
ChatGPT is the #5 website on earth with 700M+ mostly-unmonetized free users. GPT-5's router quietly classifies query intent and complexity — the one attribute needed to add commercial value. A high-intent query ('best DUI lawyer near me') could get $50 of compute because the referral is worth thousands. Fidji Simo (ex-Facebook/Instacart monetization) now runs OpenAI Applications; a take-rate/agentic-checkout SuperApp (Shopify, Instacart, Etsy already integrated) is the endgame.
“We believe that the Router is the groundwork for the next leg of ChatGPT's story, and that's monetization of free users.”— GPT-5 Set the Stage for Ad Monetization and the SuperApp (2025-08-13), SemiAnalysis
12. RL environments are spilling into the physical world — biology and materials wet-labs become the new bottleneck.
Environments are no longer just docker containers. Periodic Labs and Medra are building robotic wet-labs that generate experimentally-verified reward signals; DeepMind starts an automated materials lab in 2026. The economics invert: a coding task can be attempted 64x for trivial cost, while one biology experiment costs hundreds-to-thousands of dollars and takes hours, with sparse rewards. OpenAI targets early drug discovery (GPT-5 in a closed loop with robots); Anthropic ships Claude for Life Sciences (Benchling, PubMed connectors) for development/approval.
“RL environments are spilling into the physical world. These environments are no longer just a docker container... It is an experiment that needs to be run by a human or a robot, with a real cost to materials, electricity, equipment, and lab space.”— RL Environments and RL for Science: Data Foundries and Multi-Agent Architectures (2026-01-06), SemiAnalysis
13. RL rewires hardware and datacenter design: memory-bound, CPU-hungry, decentralizable — favoring low-$/HBM chips like Trainium.
RL uses fewer FLOPs than pre-training but heavy memory loads (rollouts, long context, judges), so the NVL72 rack-scale shared memory shines and low-TCO memory chips win. Anthropic is onboarding large Trainium volumes precisely because RL is memory-bound and Amazon sourced HBM directly to cut cost. RL environments also revived the datacenter CPU — Microsoft's 'Fairwater' pairs a 48MW CPU/storage building with a 295MW GPU cluster, and labs are now scrambling for CPU allocation.
“Anthropic is bringing on large volumes of Trainium, which is economically attractive for RL in part due to low $/HBM... RL workloads are mostly inference, which is memory bound.”— RL Environments and RL for Science: Data Foundries and Multi-Agent Architectures (2026-01-06), SemiAnalysis
14. RL-training efficiency is a queue-health problem: match trainer and generator throughput or waste your GPUs.
An open-source RL system has three actors — generator (inference), environment (sandbox), trainer. If the generator is slower the trainer starves; if faster, samples go stale. PipelineRL's in-flight weight updates let the two run at different speeds within a policy-staleness budget, ~2x more iterations per wall-clock hour. SemiAnalysis's own experiments were consistently 'generation-bound' (trainer idle 30-74%), and sandbox scaling to ~960 concurrent rollouts hit dead init errors — proving sandbox infra is a first-class bottleneck.
“RL training system efficiency is a matter of queue health... in-flight weight updates allow trainer and generator to operate at different rates.”— RL Systems Mind the Gap: Matching Trainer and Generator Throughput (2026-06-16), SemiAnalysis

Article Deep-Dives

Scaling Laws – O1 Pro Architecture, Reasoning Training Infrastructure, Orion and Claude 3.5 Opus 'Failures' (2024-12-11)
The foundational rebuttal to the 'scaling is dead' narrative. Reframes scaling as a stack: pre-training (hitting the data wall, Chinchilla-suboptimal mega-models), post-training (SFT, RLHF, RLAIF, PPO with Outcome/Process Reward Models), and test-time compute via reasoning (search, best-of-N, Monte Carlo roll-outs). The most-cited scoop: Anthropic finished Claude 3.5 Opus, it scaled fine, but they didn't ship it — instead using it to generate synthetic data and reward-model to improve 3.5 Sonnet. Details o1 vs o1 Pro (Pro adds search/self-consistency), harder evals (FrontierMath at 2%, GPQA), and why reasoning models are bottlenecked by inference systems keeping context lengths short.
Scaling Reinforcement Learning: Environments, Reward Hacking, Agents, Scaling Data (2025-06-08)
The canonical RL primer. Explains GRPO (DeepSeek's critic-free PPO variant), why RL is inference-heavy, and why reward functions are a 'dark art' prone to reward hacking (Claude 3.7 edited test files to pass). Argues 'data is the moat' since RL data must be built from scratch, and that RL fuses inference into training (labs merged inference teams). Key structural claims: RL doesn't need centralized clusters like pre-training (one lab runs it on idle inference for free compute), China is chip-constrained for the inference-heavy RL game (H20 ban bites), and recursive self-improvement is already happening via compiler/kernel RL.
RL Environments and RL for Science: Data Foundries and Multi-Agent Architectures (2026-01-06)
Maps the RL-environment industrial complex. 35+ startups sell UI-gyms (~$20k/site) and dockerized sandbox tooling (HUD wraps software + an MCP server per container); data foundries (Surge ~$1B ARR, Mercor, Handshake) recruit domain experts to author tasks and rubrics. Contrasts lab buying patterns (Anthropic = broadest vendor ecosystem + code/computer-use; OpenAI = fewer vendors but outspends + in-house Feather platform; Google = decentralized, owns the apps). Then extends to RL-for-science: Periodic Labs and Medra build robotic wet-labs, OpenAI targets drug discovery while Anthropic ships Claude for Life Sciences. Warns automation may be augmentation (GDPval: the models ran ~100x faster/cheaper but experts kept the quality edge; radiology never got automated).
Claude Code is the Inflection Point (2026-02-05)
Declares Claude Code the ChatGPT-moment for agents. ~4% of GitHub commits already Claude Code, projected 20%+ by end-2026. Frames agents as orchestration of tokens (Web 2.0) atop raw token APIs (Web 1.0/TCP-IP), and coding as the beachhead into the $15T information-work economy via the READ/THINK/WRITE/VERIFY workflow. Cowork was built by 4 engineers in 10 days, mostly by Claude Code itself. Core financial call: Anthropic now adds more revenue per quarter than OpenAI, capped by compute. Deep dive on Microsoft's conundrum — accelerating Azure means renting GPUs to the 'barbarians' (OpenAI, Anthropic) eroding seat-based Office 365; Claude for Excel is what Copilot should have been, and SaaS's 75%-margin moats are eroding.
The Coding Assistant Breakdown: More Tokens Please (2026-04-24)
A model-by-model teardown (GPT-5.5/'Spud', Opus 4.7, DeepSeek V4, Mythos, Kimi, Qwen, GLM) plus a masterclass on why benchmarks are bad. GPT-5.5 finally reaches the frontier in Codex but is beaten by Opus 4.7 on the buried Expert-SWE coding benchmark (Mythos scored 77.8%). Dissects benchmark anatomy (tasks/eval/harness) from MMLU → SWE-bench (contaminated, retired by OpenAI Feb 2026) → GDPval, and shows OpenAI's sneaky reporting. Practical workflow: engineers now start in Claude (greenfield scaffolding) then switch to Codex (harder narrow bug-fixing). The thesis pillar: cost-per-task, not cost-per-token, is the north star; the harness (Codex ~80:1 vs Claude Code ~100:1 input/output) drives the real bill.

Reference: Value Chain

Frontier closed labsOpenAI, Anthropic, Google DeepMind, xAI, Meta Superintelligence — Build and serve frontier models. Differentiation is now data procurement + post-training, not a common internet corpus: OpenAI outspends on data across many parallel programs; Anthropic is the most aggressive RL-environment buyer (first customer of dozens of vendors) and dominates agentic coding; Google is compute-rich (TPU) and owns the apps (Sheets/Docs/Drive) to post-train on.
▼
Open-weights & China labsDeepSeek, Qwen (Alibaba), Kimi (Moonshot), GLM (Z.ai), MiniMax, Meta Llama, Mistral — Commoditize the low/mid tier and set the efficiency frontier. DeepSeek's V3/V4 (MLA, MoE, FP8) and open kernels (DeepEP, DeepGEMM, FlashMLA) are used worldwide; SemiAnalysis calls DeepSeek the best open-weights lab. Still trail closed models on agentic coding and — tellingly — Claude beats them even at Chinese writing.
▼
Environments & data foundriesSurge, Mercor, Handshake, Scale AI (→Meta), Mechanize, Turing, Prime Intellect, HUD, Periodic Labs, Medra — Sell the shovels of the RL boom: cloned UI 'gyms' (~$20k/site), dockerized sandbox tooling, and expert-authored tasks/rubrics. Surge ~$1B ARR, Scale did ~$870M in 2024 (~$1.5B exit run-rate) before Meta absorbed it (and the labs fled). Periodic/Medra push environments into physical wet-labs for RL-for-science.
▼
Agents & coding assistantsClaude Code, OpenAI Codex, Cursor, GitHub Copilot, Windsurf, Cognition, Cowork — The product layer that turns tokens into outcomes. Coding is the beachhead: the six-player market is $30B+ ARR heading to $100B. Cursor/Windsurf can post-train competitive models off high-signal user data despite lacking lab resources. The harness (CLI, fast mode, 1M context, sub-agents) is now a competitive weapon.
▼
Cloud, silicon & inference stackAWS/Trainium, Azure, Google Cloud/TPU, Nvidia GB200/GB300, vLLM, SGLang, Modal — RL fused inference into training reshapes the stack: memory-bound RL favors low-$/HBM chips (Trainium), NVL72 rack-scale world size wins at reasoning, CPUs surge for RL environments, and open inference engines (vLLM, SGLang) get Day-0 support that Nvidia's own TensorRT-LLM often lags. Sandbox providers (Modal, Prime Sandbox) are new critical infra.
▼
Distribution & monetizationChatGPT (700M+ users), Microsoft M365/Copilot, Meta AI, Accenture, enterprise — Where value is captured. OpenAI's router sets up free-user monetization via agentic purchasing/take-rates; Anthropic rides enterprise + Claude Code (Accenture trained 30,000 on Claude); Microsoft is caught between renting GPUs to the 'barbarians' and defending seat-based Office 365 that agents erode.

Reference: Core Concepts

The scaling stack (pre-train → post-train → test-time). SemiAnalysis's core framing: overall AI 'scaling' is the sum of three compounding vectors, not just pre-training. Pre-training (next-token prediction on the internet) hit a data wall; post-training (SFT/RLHF/RLAIF/RL on verifiable rewards) unlocked reasoning; test-time compute lets a model 'think' longer for a better answer. Like Moore's Law surviving the end of clock-speed scaling via multi-core, the aggregate curve keeps climbing even as one vector slows.

RL on verifiable rewards (GRPO/PPO). RL worked best where a reward is cleanly checkable — code (does it pass tests?) and math (is the answer right?). The model generates many 'rollouts' (attempts) per prompt, each scored; GRPO (DeepSeek's algorithm) drops PPO's critic model to save memory by scoring each rollout against the group average. This makes RL inference-heavy: hundreds of answers per question, so production-grade inference is now inside the training loop.

RL environments & sandboxes. An 'environment' is the simulated world (a code sandbox, a cloned website, a browser, a spreadsheet app) where an agent acts and receives a reward. Engineering robust, low-latency, exploit-resistant sandboxes (Firecracker micro-VMs to full QEMU VMs) is a first-class systems problem — sandbox startup latency and scaling to thousands of concurrent rollouts are real bottlenecks. Reward hacking (the model gaming the reward without doing the task) is the recurring failure mode.

Synthetic data, RLAIF & LLM-as-judge. As human data stopped scaling, labs turned to synthetic data (rejection sampling, model-generated tasks) and replaced human feedback (RLHF) with AI feedback (RLAIF). A model or 'LLM judge' scores outputs against a rubric — the key that opens non-verifiable domains like writing, medicine and law. Better models are better judges, creating a self-improving loop. Anthropic's Constitutional AI is the canonical RLAIF example.

Agent / harness (Claude Code). An agent wraps a model in a loop of READ → THINK → WRITE → VERIFY with tools, memory and sub-agents. Claude Code is a terminal-native agent ('Claude Computer') that reads a codebase, plans, and executes. The 'harness' (tools, prompts, context management, caching) is now part of the product — it drives cost-per-task as much as the model does, which is why same-harness benchmark comparisons miss the point.

Cost-per-task vs cost-per-token. SemiAnalysis's north-star pricing metric. A model priced 5x higher per token (e.g. Mythos vs Opus) can be cheaper in practice if it solves the task in far fewer tokens ('token efficiency'). Anthropic's Opus 4.7 tokenizer change alone raised token usage up to 35% — an implicit 35% price hike. The harness (Codex ~80:1 input/output vs Claude Code ~100:1) is a major driver.

Data foundry / RL-as-a-service. Firms that supply the labs with RL environments, expert-authored tasks, and grading rubrics. Environment startups (Mechanize, Turing, Prime Intellect, HUD) clone apps and wrap software in dockerized sandboxes; data foundries (Surge, Mercor, Handshake) recruit domain experts. RL-as-a-service startups (RunRL, Osmosis, Applied Compute, Thinking Machines Tinker) post-train custom models (often Qwen) for enterprises at ~5x lower cost than OpenAI's RFT.

Multi-Head Latent Attention (MLA) & MoE. DeepSeek's signature efficiency innovations. MLA compresses the KV cache ~93% versus standard attention, slashing the memory (and thus cost) per query. Combined with fine-grained Mixture-of-Experts (many small experts, few active per token) and FP8 training, this is how a Chinese lab reached near-frontier quality at a fraction of the compute — and forced Western labs to copy the ideas almost immediately.

Open Questions

Sources (SemiAnalysis)

Independence & sourcing. This is independent analysis by Yicheng Yang, distilled from publicly accessible SemiAnalysis articles (free posts and free previews; no paywall circumvention) and verified against the underlying text. It is not affiliated with, endorsed by, or a substitute for SemiAnalysis — subscribe there for the full research. All referenced claims are sourced and linked per SemiAnalysis's attribution terms. No SemiAnalysis images are reproduced. Nothing here is investment advice.