The AI lab business: industry and inference research
Evidence cut: 2026-11-06. The dates attached to individual claims identify the underlying announcement, filing, or tariff. Announced future capacity is a promise or expectation. An activated site is a different observation. Company claims, independent measurements and the game's assumptions belong on separate cards.
The player runs one part of a value chain: accelerator makers and utilities enable physical compute; site operators and clouds sell capacity; labs turn staff, data and compute into checkpoints; serving companies deliver requests; channels bring customers. One company can sell several of those layers. The provider roster distinguishes them. The first mission uses a single thread through that chain: a cheaper enterprise quote forces a lab to change its costs before it scales.
The lab as a business
Observed as of 2026-11-06. Anthropic's Nov 2 S-1/A reported $33.242B of recognized revenue, $4.812B of functional cost of revenue, $19.240B of R&D and $6.569B of operating cash inflow for the nine months ended Sep 30, 2026. Its $18.808B of compute and infrastructure spans functions and overlaps those expense categories. It disclosed $5.0B in short-term compute for Q3, priced materially above comparable multi-year commitments. It sells metered usage and subscriptions, and some enterprise customers commit to minimum spend at a negotiated price list. This is company-level accounting. Model-level realized prices, margins, training-run costs and the terms of individual discounts remain calibration assumptions absent specific disclosures. Anthropic S-1/A, MD&A and financial statements
Game rule. Keep cash, recognized revenue, serving cost, R&D expense and future commitments in separate ledgers. Equity funding increases cash at closing and dilutes ownership. Debt adds repayments and collateral constraints. Customer prepayments move cash early, with sales recognized as usage is delivered or subscriptions elapsed. Show runway next to the payment calendar. Renewals depend on service reliability, task results and commercial terms. Paid token price, fully loaded serving cost, R&D, and contribution per successful job get separate lines. A lab can sell a profitable API while losing cash building its next model. Product options include direct API, subscriptions with usage caps, enterprise minimum-spend contracts, and cloud marketplace distribution. Keep paid API and one anchor enterprise customer in the first scenario.
Talent, data and release. Research, distributed systems, data and safeguards are distinct staffing bottlenecks. Claude Opus 4.6's historical system card describes public web information, non-public third-party data, contractor and labelling data, opt-in user data and internally generated data, with filtering, deduplication and post-training. Give datasets provenance, domain coverage, usage rights, preparation time and separate held-out evaluation sets. Research and systems teams compete for the same experiment window. Have capability trials and cost/latency checks determine product readiness, with a separate misuse and security release gate. Options include further mitigation, a monitored partner preview, and a broader rollout. Anthropic's voluntary Responsible Scaling Policy, its April 2026 limited Mythos Preview for defensive partners, and the EU's GPAI obligations for documentation and additional systemic-risk measures are separate evidence for those mechanics. Anthropic research and engineering roles · Claude Opus 4.6 system card, §1.1.1 · Anthropic on agent evals · Responsible Scaling Policy · Project Glasswing · European Commission GPAI model obligations
Who sells which layer
| Counterparty | Observed example | Playable contract |
|---|---|---|
| Hyperscaler and strategic investor | In 2025 AWS reported activating Project Rainier for Claude with nearly 500,000 Trainium2 chips. In April 2026 Anthropic announced an AWS technology spend commitment above $100B over ten years for up to 5 GW of new capacity, with about 1 GW expected by year-end 2026. AWS remained its primary training and mission-critical cloud. The announcement also described Bedrock, Vertex AI and Microsoft Foundry distribution. Amazon's announced $5B investment was included within the $65B Series H announced in May, with up to $20B additional investment described in April. AWS on activating Project Rainier · April compute agreement · Series H | Bundle silicon, staged training capacity, investment and distribution as separately priced clauses. A port to partner-specific hardware spends systems team time. Distribution brings customers with a fee and different procurement friction. Equity only enters cash on its funding milestone. |
| Neocloud | CoreWeave's 2025 10-K disclosed on-demand service alongside multi-year take-or-pay contracts averaging about five years at year-end. Committed contracts commonly included prepayments. In Q2 2026 it reported 1.5 GW active and about 3.7 GW contracted. CoreWeave 10-K · Q2 2026 results | A prepayment buys a delivery date and reserved accelerator topology. An oversized deal charges for idle capacity. On-demand covers peaks at a higher scenario price when suitable capacity is available. |
| Colocation or dedicated facility | Applied Digital disclosed 400 MW of contracted critical IT load at Polaris Forge 1 across 15-year leases to CoreWeave, with delivery phased. Anthropic's S-1/A describes its own hybrid sourcing, including leasing chips separately and hosting them in dedicated facilities operated on its behalf. Applied Digital FY2026 annual report · Anthropic S-1/A, Compute Investments | Reserve a powered site before committing to accelerators. The lab then procures equipment, fabric and operations. A delay can strand hardware and force temporary cloud spending. |
Rival behavior. OpenAI sought additional specialized compute from CoreWeave, which in September 2025 disclosed cumulative contracts worth up to about $22.4B. Its April 2026 Microsoft amendment kept Microsoft as primary cloud and retained Azure-first product shipping subject to the stated exception, while permitting sales via other cloud providers and making Microsoft's IP license nonexclusive. Model rivals as repeat buyers that can reserve the same scarce clusters, plus distinct cloud storefronts with their own commercial terms. Expose rivalry through bids, delivery queues, model launches and customer procurement. CoreWeave's September 2025 agreement · OpenAI and Microsoft amendment
The power and site layer
- A site progresses through land and permits, utility load study and service agreement, substation and cooling construction, energization, rack/fabric commissioning, and service readiness. Track grid connection date, facility MW, critical IT MW, cooling choice, local water constraints, fiber to users and internal interconnect separately. A tightly connected training reservation has topology and continuous run requirements. Regional inference has capacity and latency requirements. Earlier non-firm connections can impose curtailment. Offer the option to defer checkpointable workloads and pay for alternate serving capacity on curtailed turns. AEP Ohio's real large-load tariff sets a load ramp plus eight-year initial term and a sliding minimum billed-demand schedule. Teach this as a local tariff card. IEA, October 2026 · AEP Ohio Data Center Tariff · AWS EFA
- Power usage effectiveness (PUE) is facility electricity divided by IT electricity. Water usage effectiveness (site WUE) measures onsite water consumption per kWh of IT electricity and varies with cooling and climate. Liquid cooling at the rack can use dry or water-consuming heat rejection at the site. At a fictional steady 20 MW of IT load, PUE 1.2 implies 24 MW at the facility meter. Over 720 hours that's 17,280 MWh, or $1.0368M at an illustrative $60/MWh energy charge, before applicable demand charges, taxes and other costs. Show energy (MWh), coincident or billed demand (MW) and direct water separately. Berkeley Lab energy usage report, pp. 39–46
- Power purchase agreements have contract and delivery dates. In 2025 TVA and Kairos agreed that the future Hermes 2 plant would supply up to 50 MW to TVA's grid, with Google receiving associated clean energy attributes for regional operations. In the game, track offtake terms and attributes alongside an independent site interconnection milestone. A power contract can improve procurement and emissions accounting while the queue still governs the site's usable load. Kairos–TVA–Google announcement
Model research, serving and the price of a completed task
What the lab actually researches
| Choice | Evidence and implication for play |
|---|---|
| Start from open weights or train a base model | A licensed checkpoint lets a new lab specialize and launch sooner. Training a new base model buys control over the data and checkpoint but ties up researchers and a contiguous cluster reservation. Ai2 publishes Olmo 3’s pretraining, mid-training, long-context extension, supervised fine-tuning, preference optimization and reinforcement learning with verifiable rewards. Model each as a decision with a data prerequisite and a job in the queue. Put rights, curation, deduplication and a protected evaluation set on the data card. Ai2 Olmo 3 model flow |
| Dense vs mixture of experts (MoE) | Dense layers use their full weights while MoE layers route tokens to a subset of experts. DeepSeek-V3 reported 671B total parameters and 37B activated per token. Its report discusses load balance and cross-node communication. Give MoE separate active compute, total weight memory and interconnect requirements. Show the target hardware’s throughput and memory fit on the model card. The play advantage depends on the cluster and traffic. DeepSeek-V3 technical report |
| Small/distilled vs general purpose | A smaller checkpoint can be trained directly or distilled using outputs from a teacher. DeepSeek published R1-distilled Qwen and Llama checkpoints and identifies their base models. Give the small model strong economics in workloads it passes and a quality ceiling elsewhere. Record access and distillation terms on the card. Chinchilla studied the tradeoff between parameter count and training tokens under a training compute budget. A later study explicitly adds inference demand and finds conditions where a smaller model trained longer reduces total cost. DeepSeek-R1, Chinchilla, Beyond Chinchilla-Optimal |
| Reasoning and tool use | Qwen3’s original release includes dense and MoE checkpoints with selectable thinking and non-thinking modes. R1 demonstrates reasoning-oriented post-training. OpenAI documents invisible reasoning tokens billed as output on supported models. Make reasoning capability a training choice and reasoning effort a serving policy. Let hard jobs benefit from extra effort where measured, while latency, token use and possible truncation change. Qwen3, DeepSeek-R1 paper, OpenAI reasoning guide |
| Specialist workload or new generation method | Classification, embedding/retrieval and reranking are useful specialist roles if “decision” means choosing an action. If the intended term was “diffusion,” LLaDA is a documented example of masked diffusion for language generation. Keep both as future branches. The first scenario needs one small dense model, one more capable reference model and an effort setting. FlagEmbedding, Large Language Diffusion Models |
Keep separate estimates for an individual run, the full R&D program and ongoing serving. DeepSeek-V3 reported 2.788 million H800 GPU-hours across its official training stages. Its $5.576 million figure applies an assumed $2 per H800 GPU-hour, and its accounting excludes earlier research and ablations. That supports an evidence card for a run scope. Staff, data, unsuccessful experiments and deployment have separate budget lines. DeepSeek-V3 report, Table 1, Cottier et al. cost methodology
What drives serving cost and demand
- Prefill, decode and deadlines. Input length and repeated prefixes shape prefill work. Output length, including any billed reasoning, shapes decode work. Track p95 time to first token for chat, inter-token delay and successful jobs within the promised deadline. An API may return a first token quickly while the user waits on the full answer. DistServe uses time to first token and time per output token to measure goodput under SLOs. DistServe, OSDI 2024, MLCommons workload metrics
- Cache and batch. A repeated eligible prefix can reuse the KV cache, saving prefill. A one-off prompt sees little benefit. Cache writes and reads may be billed differently. Continuous batching schedules concurrent requests inside the server and changes the throughput and latency curve. An asynchronous batch product lets customers trade a deadline measured in hours for a discount. Expose both levers independently. vLLM prefix caching, vLLM scheduler optimization, Claude batch processing
- Compression, routing and utilization. Quantized weights can make a model fit in less memory. Measure resulting throughput and held-out quality by model, format and hardware. Route simple requests to the small model and escalate the hard ones. False confidence, retries and escalations belong on the bill, and excessive queueing erodes goodput. A capacity block paid for throughout the month costs more per job when demand is low. Keep a workload-specific quality floor so aggressive routing or compression can lose a contract. AWQ paper, RouteLLM code and experiments, DistServe
- Three columns in the ledger. Show retail bill, internal serving expense and contribution. Retail bill sums the disjoint metered token classes times their rates per million, plus metered tools and applicable adjustments. Serving expense tracks purchased API, rented accelerators, or owned assets with allocated depreciation and power as applicable, plus operations and retries. Divide all serving expense in the cohort by jobs that meet quality and deadline to expose cost per successful job. Zero successful jobs displays “no jobs passed” alongside the spend. For a rate comparison, use the same requests and count the tokens per model. Tokenizers can segment the same text differently. Anthropic pricing and tokenizer notes, OpenAI usage fields
- Pinned illustration. Anthropic’s May 27, 2026 global standard Claude API table lists Sonnet 4.6 at $3 input, $15 output, $3.75 per 5-minute cache-write class and $0.30 per cache-hit class, all USD per million tokens. Under that tariff, an illustrative 10,000-token prefix and 1,000 billed output tokens cost $0.045 without caching, $0.0525 on a fully eligible first write and $0.018 on a full hit. A changed prefix, expiry, or longer response changes the bill. These calculations describe customer charges under a dated tariff. Anthropic dated price sheet, Claude API section
- Price pressure can come from a launch. Anthropic's Sonnet 5 announcement says its $2 input and $10 output per million token introductory prices became permanent. It also notes a tokenizer change, so a fixed text task can use a different number of tokens. DeepSeek announced an API price reduction with V4.1 Flash on September 10, 2026. In the game, a buyer can reopen a deal when a new version meets its workload at a better effective price. Compare each model's actual request bill and success rate; do not apply one fixed annual price decline to all rivals. Sonnet 5 announcement · DeepSeek change log
Use a protected, versioned task suite for simulated contract acceptance, e.g. extraction validity and exact-field accuracy for batch, and issue resolution plus response deadlines for support. Keep capability, safety and performance results visible as separate dimensions. Label simulated scores and compare reruns under the same workload mix. Stanford HELM, MLPerf Inference Datacenter
What to calibrate
| Evidence we can pin | What it anchors | What remains a game assumption |
|---|---|---|
| Dated direct rate card | Input, output, cache, region, batch and tier categories for a customer bill | Wholesale economics, margins, actual discounts and demand elasticity |
| Workload test with model revision, hardware, precision and traffic | Serving quality, successful throughput and percentile latency within that test | Values outside the measured profile; no conversion from an API list rate |
| Company filing or dated deal announcement | Size and structure of commitments, categories of cash and spending, time to delivery | How much a given startup pays per GPU hour, when its site opens, and its financing odds |
| Utility tariff, interconnect document or power announcement | Billed MW, delivered MWh, curtailment conditions, delivery and credit structure | Synthetic region, permits, water budget and outage events for the game |
For v0, keep three distinct quantities on every turn: customer tariff, expense of serving the cohort, and cash due even if idle. Calculate cost per successful job after paying for all attempts and commitments that served the cohort; compare only jobs of the same kind with the same deadline. Put pretraining and post-training experiments in R&D. Show site capex in the cash forecast and allocation/depreciation in an optional accounting view; combining both in one period expense would count the same asset twice. Treat vendor credits as temporary and carry the expiry. For the first mission, make the assumptions explicit and fixed. A future calibration pass can substitute one measured workload and compare player choices without rewriting the rest of the economy.
For reading lists, simulators, and comparable games, start with the GitHub and precedent shortlist.