How AI labs build, sell, and fund products
Evidence checked November 6, 2026. Lab Builder draws on public product documentation, announcements, and filings. Deal amounts, prices, demand, timing, eval scores, and cost multipliers in the game are simulated. Follow the links on a card for the real example behind it.
What the labs ship and sell
Product snapshot checked November 6, 2026. The company and model names below are evidence for fictional product cards. Pricing and capacity terms on the game cards are scenario assumptions.
| Layer | Anthropic example | OpenAI example | Decision for the player |
|---|---|---|---|
| Model portfolio | Haiku 5.5 handles high-volume workloads, Sonnet 5.5 balances speed and capability, Opus 5.5 targets long-running coding and knowledge work, and Fable 5.1 targets demanding reasoning. Claude model catalog | GPT-6 Luna targets focused high-volume workloads, GPT-6.1 Sol balances capability and cost, GPT-6.1 Terra focuses on fast visual reasoning and computer use, and GPT-6 Astra targets demanding reasoning and coding. OpenAI model catalog | Train or adapt a checkpoint and choose a position in the portfolio. Evaluate each model on the workload it will serve. |
| End-user products | Claude is the general assistant. Claude Code is a coding agent that works from the terminal, IDE, web and other clients. Claude Code can draw on plan entitlements or metered usage through a Console account. Claude Code, Console billing | ChatGPT is the general assistant. Codex adds coding and review workflows through desktop, CLI, IDE and web. Hosted tasks have plan eligibility and workspace controls. Desktop, CLI and IDE support API key billing for local workflows. Codex plan access, Codex authentication | A coding product requires repository context, a tool harness, execution environments and verified patch results. It can use a model the lab already serves. |
| Commercial channels | Individual and organization plans package access to Claude products. The Claude API meters usage, commonly against purchased credits or monthly invoices under an invoicing arrangement. Eligible Max, Ultra and Team plans can claim API credits. Partners also serve select Claude models. Claude plans, API billing, Cloud platforms | Individual and business plans bundle different access and limits. ChatGPT and the API platform have separate billing, with eligible Pro and Business plans offering API credits for a linked organization. ChatGPT plans, API billing, API credit benefit | Keep subscription revenue, metered developer API revenue and enterprise agreements in separate ledger lines. Forecast their serving costs from actual demand. |
Enterprise capacity. OpenAI Reserved Tier sells model-specific provisioned throughput measured in USD per minute, with usage above reservation billed on demand. Its legacy Scale Tier uses distinct input and output token units with a minimum 30-day term and covers older releases. Anthropic’s public Priority Tier page now refers to existing commitments and directs new guaranteed-capacity buyers to sales. The game can offer a fictional reservation contract with throughput, term, guaranteed capacity and an overage rate. OpenAI Reserved Tier, OpenAI Scale Tier, Claude service tiers
Vision, image and video. Current Claude models take text and image inputs and produce text. Claude can generate diagrams through HTML/SVG. Photo and illustration synthesis belongs on a separate model card. OpenAI launched GPT-Image-2.6 Flare and Sunburst on November 4 for native generation and editing through the API, with Images 2.6 also available in ChatGPT and Codex. Sora’s web and app experiences ended April 26, 2026. OpenAI removed its Videos API and Sora 2 models on September 24, 2026. Use Sora as a dated historical case. Google’s Veo 3.1 provides an active video example billed per generated second. Claude model inputs and outputs, Claude image help, OpenAI Images 2.6, Sora discontinuation, OpenAI API removals, Veo pricing
Architecture evidence. OpenAI’s gpt-oss-120b is a released open-weight MoE with about 117 billion total and 5.1 billion active parameters per token. Track both total weight memory and per-token work. OpenAI’s original Sora report describes diffusion over video and image latent patches using a transformer, which supports a distinct video research branch with temporal coherence, clip duration and compute per render. Treat these as facts about the named releases and use the current model catalog to place commercial products. gpt-oss release, Original Sora technical report
Named deal cards and market behavior
Evidence checked 2026-11-06. Keep announced ceiling, binding spend, closed investment, undrawn credit, scheduled capacity, accepted capacity, recognized revenue and collected cash as separate states.
| Actors and roles | Public evidence | Decision this supports |
|---|---|---|
| Amazon / AWS ↔ Anthropic. Strategic equity investor, cloud and chip supplier, and Bedrock distribution channel. | April 20, 2026 announcement expanded the AWS technology commitment by more than $100B over ten years for up to 5 GW of new compute, with nearly 1 GW expected by year-end 2026. Amazon's Sep. 30 10-Q separately reports $5B Series G and $5B Series H nonvoting preferred investments in Q2. The Series H purchase reduced the conditional $20B financing facility to $15B. Availability depends on capacity delivery milestones. These are distinct capital flows and future obligations. Anthropic announcement · Amazon Q3 2026 10-Q, notes 1 and 2 | Negotiate spend floor, performance acceptance, silicon port and staged delivery independently of equity and marketplace reach. Financing that unlocks when capacity is delivered cannot fund a bridge before the milestone. |
| Amazon / AWS ↔ OpenAI. A supplier and investor can serve rival labs. | February 27, 2026 announcement added $100B over eight years to a prior $38B agreement, including about 2 GW of Trainium capacity. Amazon's Q3 10-Q reports the $50B Series C investment had been made by Sep. 30. OpenAI confirmed June 1 availability of models through Bedrock and Codex using AWS credentials and infrastructure. Those launches and the investment are observed events, while the Trainium commitment refers to future consumption. February agreement · June launch · Amazon Q3 10-Q | A rival can win the same provider's accelerator queue. Separate distribution access by product and marketplace from the site supplying inference or training. |
| Microsoft / Azure ↔ OpenAI. Equity investor, primary cloud, IP licensee, and revenue-share recipient. | Microsoft reported about 24% as-converted ownership, $13B funding committed and $12B funded as of Sep. 30, 2026. It recorded $10B in Q3 commercial revenue from OpenAI including revenue sharing and reported $7.2B of quarter-end accounts receivable. The April 2026 amendment retained Azure-first launches subject to the stated exception and allowed service through other clouds. Microsoft's IP license is nonexclusive through 2032. OpenAI's capped revenue share to Microsoft continues through 2030. The October 2025 announcement's roughly 27% and $135B represented then-current stake ownership and value at recapitalization. Microsoft Q1 FY2027 10-Q, related-party disclosures · April amendment · October recapitalization | A cloud anchor can ask for first-launch priority and a time-limited share of receipts in exchange for funding and distribution. Draw a distinction between ownership value, cash disbursed, earnings and cash collected. |
| Meta ↔ Scale AI, open weights and infrastructure investors. The rival can fund its data supply, distribute models and finance sites through different arrangements. | Meta's 10-Q says it completed a nonvoting minority Scale AI investment in June 2025, with $13.79B of the consideration recorded as nonmarketable equity. Meta also publishes Llama 4 Scout weights under a custom community license and launched Muse Wonder via Meta Model API and Muse Code in October 2026. Its October 2025 Hyperion JV with Blue Owl funds split ownership 20% Meta / 80% fund and committed pro rata to about $27B in estimated development cost. Blue Owl funds contributed about $7B in cash at formation. Meta took an initial four-year lease plus renewal options and gave a capped conditional residual-value guarantee. Meta Scale investment filing · Llama model card and license · Muse Wonder launch · Hyperion JV terms | Licensed weights increase the customer's options and can press prices for routine work. A JV brings site capital forward and leaves the lab with rent and potential guarantee exposure. Equity ownership of a supplier and access to its data can be separate clauses. |
| SpaceX / xAI ↔ Anthropic. An AI rival's parent can also rent compute. | SpaceX's Sep. 30 10-Q confirms its completed xAI acquisition on Feb. 2, 2026. In May Anthropic announced access to all Colossus 1 compute. SpaceX's June final prospectus describes the May contracts spanning Colossus and Colossus II, covering about 325,000 NVIDIA GPUs, $1.25B/month through May 2029, and discounted May/June ramp fees. Either party can terminate on 90 days' notice after the first three months. SpaceX's September filing reports cloud-services results in aggregate. May Anthropic announcement · SpaceX June final prospectus, Cloud Services Agreements · SpaceX Q3 10-Q | Buy rapid bridge capacity while preparing a fallback for the notice window. Give the supplier separate decisions about renting capacity to rivals versus preserving it for its own model. |
| Barclays ↔ Anthropic. Regulated enterprise rollout. | The Oct. 1, 2026 announcement describes expanded enterprise use, with a target of Claude Code adoption by 50% of Barclays developers by year-end. It provides adoption targets and use cases. For a disclosed commercial discount, Anthropic's Oct. 27 GSA agreement announced 40% for eligible customers, available direct at launch and with cloud rollout planned later. Barclays announcement · GSA announcement | Model compliance review, a pilot, measured seat adoption, a wider rollout and renewal. A discounted channel can improve access while lowering yield. Track announced adoption targets separately from active seats, consumed usage and revenue. |
Optional neocloud card. Nscale's September 2026 filing discloses up to about $44.6B of Anthropic GPU services payments at Monarch Compute Campus, contingent on delivery and service availability. It described four NVIDIA Vera Rubin NVL72 tranches, each with a term starting on acceptance. At the filing date it reported financing still in progress. Simulate a signed deal whose capacity remains zero until financing closes, construction and accelerator delivery finish, and the customer accepts a tranche. Nscale prospectus, Customers and Business Model
What a recipe changes
| Choice | Real-world meaning | Lab Builder |
|---|---|---|
| Architecture and checkpoint | Dense language models use their full weights. In a mixture of experts, routing reduces active work per token while total weight memory and expert communication still matter. A licensed checkpoint can be fine-tuned, and a distilled student learns from eligible teacher outputs. OpenAI gpt-oss release · DeepSeek-R1 model releases | Choose language, image, or video, then train from scratch, adapt open weights, or distill a dense student. Checkpoint provenance and benchmark coverage affect what you can serve. |
| Training and serving stack | PyTorch and CUDA are one route. AWS Neuron documents PyTorch training on Trainium and vLLM serving for supported language checkpoints. Google Cloud documents TPU inference through its JAX vLLM plugin. Diffusers offers pipelines for image and video generation. AWS Neuron · Google Cloud TPU · Diffusers | The CUDA profile covers text and media. Trainium and TPU are qualified for the scenario's text checkpoints and carry setup costs. A lease offsets eligible compute only during its term. An active training run pays its budget at booking. |
| Data, optimization, and release | Licensed examples and held-out evaluations connect training to a specific workload. Compression, batching, and caching change memory, throughput, or latency and should be validated against that workload. vLLM documentation · Anthropic on agent evaluations | Better data can clear a demanding gate. Quantization lowers the modeled serving cost and reduces the modeled score. Media batching lowers cost and reduces demand from users who want faster renders. A checkpoint needs its own release evaluation. |
How to read the ledger
Model budgets include run compute, data, stack setup, and serving preparation. Evaluation is booked separately. The plan shows a checkpoint date and the earliest quarter the model can serve. A coding agent adds tool and sandbox costs to a qualified language model. Media has separate bills per image and per generated second.
Equity adds cash and changes founder ownership. A reservation charges a minimum across a specific term. A customer advance enters cash when collected and offsets that customer's invoices as service is delivered. The recurring operating result uses earned product revenue less inference, tools, distribution shares, staff and recurring capacity payments. Closing cash also includes financing, upfront costs, and payment timing.