The Economy of Tokens

Below top-k · 25

• claimThe open frontier is driven by recipe diffusion, not distillationc 0.85

  The popular view that open models are cheap tracings of closed ones is mistaken. What drives the open frontier is the diffusion of architectures and training methods that move freely between labs.

• claimThe closed-model premium collapses once open is 'good enough'c 0.85

  Closed frontier labs hold a differentiation premium only as long as the quality gap is real. The moment open weights cross the good-enough line for a workload, the premium on that workload collapses toward zero — the same Bertrand-style dynamic that hollowed out Sun Microsystems.

• implicationWhether intelligence becomes a commons or a chokepoint is undecidedc 0.85

  The disaggregation is real and economics favor it, but strategic markets this large rarely settle unaided. The outcome will be decided less by what the models can do than by who is permitted to build them.

• claimSoftware engineering is now a property of the wider frontierc 0.75

  Open-weight models like Composer, GLM, Kimi, Nemotron, and Minimax are widely deployed as software engineers, and a substantial share of the 400T tokens served is software. The task is being solved by AI as a category, not by an exclusive frontier oligopoly.

• contextModular architectures beat vertical integration in tech industriesc 0.70

  Baldwin and Clark argued that the pivotal economic event in technology is often not a new product but the creation of a modular architecture with stable interfaces. The PC ecosystem out-innovated vertically integrated computer companies because Intel, Microsoft, and thousands of independent vendors could each improve their layer.

• mechanismThe transformer learns structure itself rather than domain priorsc 0.70

  Earlier neural networks exploited domain-specific structure like locality in images or sequentiality in language, but the transformer learns structure from token sequences directly. This is why the same architecture models language, code, images, proteins, and physical systems.

• evidenceModern transformer blocks are assemblies of borrowed partsc 0.70

  Rotary embeddings, RMSNorm, SwiGLU, grouped-query attention, and mixture-of-experts each originated in specific papers and became near-universal defaults across Minimax, Mistral, Qwen, and DeepSeek. Open models today look nearly identical under the hood.

• evidenceInfrastructure capital is disaggregating into tradeable piecesc 0.70

  CoreWeave raised $8.5B lent against the GPUs themselves, Meta's Hyperion campus is financed by Blue Owl, and sovereign wealth funds like the UAE's $100B vehicle and Saudi Arabia's HUMAIN pay for capacity as critical infrastructure. Buildings, power, chips, and financing are coming apart into separate markets with their own specialists.

• implicationOpen weights make GPU clusters into securitizable assetsc 0.70

  A cluster financed against one company's closed model is a bet on that company; the same cluster serving open weights is a general-purpose asset, fungible across customers and models. Fungibility is what lets capital securitize the layer at scale.

• evidenceSWE-bench went from 2% to over 85% in three yearsc 0.65

  Frontier models scored around 2% on SWE-bench at publication and most models now sit above 85%. The inevitability of software AGI, however defined, is now a mechanical extrapolation rather than an act of imagination.

• evidenceOpen-weights labs are building real revenue enginesc 0.65

  Kimi's ARR crossed $100M in March 2026 and doubled past $200M within a month, and MiniMax is showing a similarly rapid ramp. Open labs monetize by selling fully produced tokens and by licensing weights to distributors, telcos, and companies building their own IP.

• evidenceToken volumes grew 10,000× in nine monthsc 0.60

  At Together Compute, tokens processed through APIs rose from 30B to 400T per month in nine months, with autonomous agents emerging as tokens became dramatically more useful.

• implicationTransformer standardization created a stable target for hardwarec 0.60

  The uniformity of transformers gave NVIDIA a decade-long roadmap and let AMD, Google TPU, AWS Trainium, and Cerebras compete by overfitting to a known computational profile. New silicon entrants only need to claim a place on the price-performance Pareto rather than convince the world it needs a new chip.

• evidenceTraining methods diffuse across labs within weeksc 0.60

  RLHF, DPO, Constitutional AI, and DeepSeek's GRPO with the R1 long-reasoning recipe each spread across the ecosystem shortly after publication. Isolated breakthroughs become industry-wide productivity gains atop the shared platform.

• contextMCP became the standard interface between harnesses and toolsc 0.60

  Anthropic open-sourced the Model Context Protocol in late 2024, and within eighteen months it became the shared standard adopted by major model labs and tool companies. A tool written once works with any harness, and models can be swapped without rewriting anything above them.

• evidenceVenture is pricing open weights as a strategic assetc 0.60

  Mistral raised €1.7B at €11.7B, Moonshot went from $4.3B to $20B in months, Zhipu and MiniMax listed publicly, and Reflection AI raised $2B to be the Western open-frontier lab. The four front-rank Chinese open labs together carry more than $180B in value.

• mechanismEngine-agnostic kernel libraries let optimizations diffuse in daysc 0.55

  Because FlashAttention, CUTLASS, and similar kernel libraries work across every inference engine, an improvement by one author propagates to the entire industry within days of release.

• mechanismSoftware already wraps most of life in callable toolsc 0.55

  Calendars, inboxes, payment rails, documents, and booking systems are all tools a harness can call, so the act-observe-correct loop reaches across most of digital life. Coding sharpened tool use generally, letting the harness pattern jump quickly from IDEs to consumer contexts.

• mechanismCoordination interfaces have become training targetsc 0.55

  Frontier labs now post-train models to operate inside harnesses using RL against live environments, with process rewards grading every tool call and step of a long trajectory. MiniMax trained M2.5 across more than a hundred thousand real agent environments this way.

• mechanismNVIDIA commoditizes the complement with Nemotronc 0.55

  NVIDIA gives away weights, datasets, post-training recipes, and RL environments because its largest customers are also its emerging rivals designing their own accelerators. A broad population of open models runnable on anyone's hardware prevents the market from consolidating onto silicon controlled by a few frontier companies.

• caveatA wholesale Sun-style collapse of closed labs is unlikelyc 0.55

  The author does not expect closed frontier labs to be wiped out, but market formation around open weights is already impossible to stop. Open will keep commoditizing workloads that become popular and application-independent.

• contextTokens are becoming a fundamental resource like electricityc 0.55

  The world runs 15–20 GW of AI capacity today, and McKinsey and Goldman expect 150 GW within five years, implying $6–7 trillion of infrastructure investment. Producing tokens takes flops, and flops take power.

• mechanismNeoclouds arise because transformer workloads are fungiblec 0.55

  Because every transformer runs on the same GPU cluster, capacity underwritten for one customer is readily resold to the next. Standardized, thin workloads make specialist neoclouds far cheaper to stand up than a full hyperscaler.

• implicationSovereign states will demand independence in intelligence productionc 0.50

  Nations pursued energy independence and will similarly refuse to let a foreign stack supply a strategic input like intelligence. Building the frontier will only get easier as the substrate matures and each layer gets cheaper to build on.

• exampleSlime shows RL post-training as modular compositionc 0.40

  Zai's slime framework, used to post-train GLM-5.2, bridges Megatron's training loop and SGLang's rollout engine through a shared data buffer without wrapping them in new abstractions. Advances in one component diffuse to the rest within a release cycle.

Redundant with selected · 3

• claimAI has converged on three standardized interfacesc 0.95 · sim 0.86

  The transformer architecture, the OpenAI-compatible inference API, and agentic harnesses have quietly standardized the production, consumption, and coordination of intelligence, driving progressive unbundling of the stack.

  overlapped with: The same modularization is now happening in AI

• claimThe transformer has become the industrial substrate for intelligencec 0.85 · sim 0.82

  Like steam engines and shipping containers before it, the transformer turned fragmented craft into an industrial system. Every improvement to attention, optimizers, kernels, or frameworks now advances almost every model at once.

  overlapped with: Open-weights models are an economic consequence of modularization

• exampleMiniMax M2.7 was 5× faster and 63× cheaper than Opus 4.8c 0.50 · sim 0.82

  Head-to-head agentic tests across MiniMax, Kimi, and Opus showed dramatic cost and speed advantages for open models, with accuracy tradeoffs closing rapidly. Companies like Coinbase are already mixing closed and open tokens through harnesses for optimal cost.

  overlapped with: Open-weights tokens are categorically cheaper, not just marginally