OpenAI

San Francisco AI lab that turned ChatGPT into the world's first consumer AI brand and the largest closed-frontier model business by revenue.

1. Core Product / Service

OpenAI ships three product surfaces stacked on top of one model line:

• API platform (platform.openai.com) — chat completions, responses, batch, fine-tuning, embeddings, image / audio / video. OpenAI-compatible endpoint format that the rest of the industry standardized around.
• ChatGPT — consumer + Team + Enterprise + Edu. The single largest paid AI consumer product (Plus $20/mo, Pro $200/mo, Team / Enterprise seat-based).
• Stargate infrastructure JV — multi-hundred-billion-dollar US data center buildout with Oracle, SoftBank, Microsoft (see §4).

Model line (as of 2026-05): GPT-5 family (released 2025) is the current default; GPT-5 Pro / GPT-5 Thinking are the high-end reasoning SKUs; o-series reasoning models continue as the test-time-compute lineage. GPT-6 has been previewed publicly but is not yet the default API SKU [1].

2. Target Users & Pain Points

• Consumers — ChatGPT free + Plus dominate B2C, the only AI product with measurable cultural ubiquity.
• Enterprise / dev — API + ChatGPT Enterprise serve the largest closed-model enterprise pipeline in the industry; Microsoft / Azure OpenAI Service distributes the same models to Fortune 500 procurement.
• Agents / advanced devs — Operator, Codex, Atlas (browser agent), Deep Research, GPT-5 Pro reasoning are the surfaces for high-value agentic work.

Pain solved: highest-quality general-purpose model, broadest tool ecosystem, deepest enterprise distribution. The premium price is the trade.

3. Competitive Landscape

OpenAI's defensive position: ChatGPT brand + the deepest consumer + enterprise distribution stack. Vulnerability: open-weight Chinese labs (DeepSeek, Kimi, Qwen, GLM) reset the price floor repeatedly, forcing OpenAI's API margins down even as ChatGPT-side margins stay healthy.

4. Unique Observations

• Frontier training cost (GPT-5 family): OpenAI does not disclose training cost. Independent estimates of GPT-5-class training runs sit in the $0.5B–$3B all-in range (compute + data + people + failed runs amortized), versus GPT-4's roughly $80M–$100M compute-only at 2023 prices [1]. The order-of-magnitude jump is the qualitative point — multi-billion-dollar single-model training is the new frontier benchmark.

• API pricing — top SKU (GPT-5): published list price for the GPT-5 default tier is $1.25/M input · $10/M output (as of 2026-05) [2]. GPT-5 Pro / Thinking SKU is materially higher (multi-dollar input, $30+/M output). Compared to deepseek V4 (input $0.27/M cache-miss, output $1.10/M), OpenAI is ~5–10× more expensive on input, ~9× on output for the standard SKU.

• Pricing vs estimated unit cost — gross margin signal: With H100/H200 inference economics, the per-million-token marginal cost on a GPT-5-class MoE is plausibly in the $0.10–$0.40 range output, implying API gross margin >90% at list price on cache-warm workloads. The actual blended margin is lower because (a) ChatGPT free users carry inference cost with no direct revenue, and (b) Stargate-scale capex amortization will eventually be folded back into unit-cost accounting. Reuters / The Information have repeatedly reported OpenAI is gross-margin positive but operating-margin deeply negative through 2026 because of training + capex.

• Open vs closed: closed weights, with one notable exception — OpenAI shipped an open-weight reasoning model (gpt-oss) in 2025, framed as a research / safety release rather than a product. Strategic logic: keep the moat on the production frontier, give the open ecosystem a credible breadcrumb so Chinese labs do not own the open-weight narrative entirely. The default for every commercially relevant model remains closed.

• Vertical integration (Stargate): announced 2025-01 by Trump + Altman + Son, Stargate is a $500B (announced) US AI infrastructure JV — OpenAI + Oracle + SoftBank, with Microsoft as anchor cloud customer. First Abilene, TX site under construction. This is the single most aggressive L3 → L1 vertical play in the industry: OpenAI bypasses coreweave / nebius / hyperscaler GPU markup entirely on its own training and serving capacity. Microsoft remains the privileged cloud distribution partner via Azure OpenAI Service [4][5]. Compare the scale: OpenAI Stargate ($500B over 4 years) vs xAI Colossus (~$10B-class) vs Anthropic's AWS Trainium commitment ($4B+).

• Revenue run-rate / consumer split: 2026 reporting puts OpenAI's annualized revenue in the $20B+ run-rate range, with the majority from ChatGPT subscriptions rather than API — the largest single shift since 2023, when API was the dominant share. ChatGPT consumer is the cash cow funding Stargate capex. Headcount has crossed 3,000+ employees (vs ~770 at end of 2023) [3][6].

• Backlog vs cash reserves — structural funding gap: OpenAI's aggregate signed AI infrastructure commitments (Azure, Oracle, AWS, CoreWeave, Stargate JV) exceed its current cash reserves by a material margin. Unlike anthropic, whose Series H ($65B round) saw broad cloud-provider participation with aligned incentives, OpenAI's financing structure embeds tension between cloud providers as investors, creditors, and competitors [7]. The practical consequence: without a successful IPO in the 2026–2027 window that raises sufficient primary capital, OpenAI cannot fully cover its signed commitments from existing balance-sheet resources [7].

• IPO risk thesis: Internal analysis (2026-06-27) predicts OpenAI's IPO will either fail to materialize in 2026–2027 or price at a valuation well below its last private round. Three structural headwinds: (1) funding structure carries too many external obligations that depress free cash flow available to public shareholders, (2) revenue growth faces competition from Gemini on the consumer side and from anthropic plus Chinese open-weight labs (deepseek, kimi, qwen) on the enterprise/API side, (3) model training costs continue rising per generation while marginal capability gains diminish. Quantitative checkpoint: valuation below $800B at 2026-12-31 as the benchmark for a disappointing outcome [7].

5. Financials / Funding

• Revenue (annualized run-rate): ~$10B in 2024 → reported step-up well into the tens of billions by 2026, weighted to ChatGPT [3].
• Headcount: ~3,000+ as of early 2026.
• Profitability: gross-margin positive, operating-margin negative due to training + Stargate capex.
• Comparison note: anthropic closed its Series H at $65B in 2026, with all major cloud providers participating — a structurally different (and arguably healthier) financing pattern vs OpenAI's cloud-provider tension model [7].
• Signed infra commitments: material multi-hundred-billion-dollar aggregate across Azure, Oracle, AWS, CoreWeave, and Stargate — some with binding take-or-pay terms, others softer. The binding vs non-binding split is not fully public but the aggregate exceeds current cash reserves [7].

6. People & Relationships

• CEO: Sam Altman.
• President / co-founder: Greg Brockman.
• Chief Scientist (post-Sutskever): Jakub Pachocki.
• Notable alumni: Ilya Sutskever (now SSI), Mira Murati (now Thinking Machines), Dario + Daniela Amodei + most of the anthropic founding team.
• Investors: Microsoft (largest), Thrive Capital, SoftBank, Khosla, Tiger, Sequoia, Founders Fund, plus sovereign-wealth participation.
• Cloud / infra partners: Microsoft Azure (anchor), Oracle (Stargate), SoftBank (Stargate), CoreWeave (capacity contract).
• Competitors: anthropic, google-deepmind, xai, deepseek, kimi.