Research OpenAI's disclosed or publicly estimated operating costs as of mid-2026, including compute/inference spend…

OpenAI’s cost structure remains dominated by inference and training compute on Microsoft Azure (plus growing multi-cloud commitments), with equity-heavy compensation amplifying the burn even as gross margins improve modestly from scale and efficiency gains. As of mid-2026, the company is on track for roughly $14 billion in 2026 losses (internal forecasts), with Q1 2026 alone showing $5.7 billion revenue, $3.7 billion cash burn, and 39% gross margins (up from 33% in 2025).[1][2][3]

Inference costs alone hit $8.4 billion in 2025 and are projected at $14.1 billion in 2026; total 2025 expenses reached $34 billion against $13.07 billion revenue (operating loss ~$20.9 billion before one-time items).[4][3] Azure commitments include an incremental $250 billion (part of a ~$281 billion Azure backlog share), with flexibility for other providers like Oracle/CoreWeave but continued heavy reliance on Microsoft infrastructure.[5][6]

Headcount has scaled rapidly (estimates range ~4,000–7,850 by late 2025, targeting ~8,000 by end-2026), with average stock-based compensation of ~$1.5 million per employee in 2025 (for the ~4,000-employee base), making equity grants a multi-billion-dollar annual expense.[7][8][9]

Gross margins vary significantly by segment, with API and enterprise deals likely achieving higher margins than high-usage ChatGPT consumer subscriptions (which carry free-tier drag and intensive inference). Overall gross margins sit at 33–39%, constrained by inference as the primary cost of revenue; “compute margins” (revenue after direct inference) have improved sharply from ~35% in early 2024 to ~70% by late 2025 due to optimizations, cheaper hardware, and model efficiency.[10][3]

Enterprise now exceeds 40% of revenue (on track for parity with consumer by end-2026), while API/licensing is estimated at 15–20%. Consumer subscriptions (ChatGPT Plus/Team) and free-tier usage drive the bulk of inference volume but face lower effective pricing per query and higher per-user costs; enterprise/API benefit from committed spend, customization, and potentially better utilization or caching. Exact per-segment breakdowns are not publicly disclosed, but inference remains the shared margin pressure point.[3][11]

Inference cost per query/token has declined dramatically (roughly 10x annually in equivalent capability tiers), but total spend continues rising with usage scale and more complex reasoning models. GPT-4-class performance fell from ~$20 per million tokens in late 2022 to ~$0.40 by 2025/early 2026, enabled by hardware price drops (H100s stabilized lower), quantization, speculative decoding, smaller/distilled models, and competitive pressure (e.g., DeepSeek claims 20–50x cheaper inference).[12][13]

Per-query costs historically ranged from single-digit cents (simple prompts) to higher for long-context/reasoning; trends point to further compression (potentially $0.10–0.20 per today’s $1 query in 18 months), though frontier models retain premium pricing and reasoning chains inflate token counts. OpenAI’s own API pricing reflects this (e.g., flagship models in the $5+/M input, $30+/M output range as of mid-2026), with caching and provisioned throughput offering savings.[14][15]

Analyst and internal forecasts point to breakeven or cash-flow positivity only in 2029–2030 (or later in the 2030s per some scenarios), after cumulative losses/burn of ~$115 billion through 2029, with 2026 losses projected at ~$14 billion (some estimates $17 billion+).[16][3][17] The Information and WSJ reporting, along with Sacra analysis, highlight that revenue growth (targeting ~$20–30 billion annualized by end-2025/2026) is outpaced by compute, R&D, and hiring; structural profitability requires sustained margin expansion, mix shift to enterprise, and slower model refresh cycles.[1][18]

Cost Structure Breakdown (Approximate/Synthesized 2025–2026 Annualized View)

Category	Estimated Annual Cost (USD)	Notes / % of Total
Inference/Compute (COGS)	$8.4B (2025) → $14.1B (2026 proj.)	Primary gross margin driver; Azure-heavy with multi-cloud diversification.[3]
R&D (incl. training compute, salaries, data)	Major share of remaining opex (~$10B+ combined with other)	Includes frontier model training; one estimate ~$8.3B AI compute + $1B data in 2025.[19]
Compensation (cash + equity)	~$6B+ equity alone (at 4k employees × $1.5M avg); total comp higher with headcount growth	Equity ~46% of 2025 revenue in some analyses; median total comp ~$600k–$800k.[7][20]
Sales & Marketing	~$5.7B (2025 example)	Aggressive enterprise push.[4]
G&A + Other	~$1.6B+ (2025 example)	Includes facilities, legal, etc.[4]
Total Expenses	~$34B (2025 reported); higher 2026 run-rate	Leading to $14B–$27B burn projections.[4][3]

Capital intensity is extreme in the current phase—hundreds of billions in multi-year cloud commitments plus continuous model training—but could moderate at scale if per-token costs keep falling, utilization improves, and revenue mix shifts toward higher-margin enterprise/API. Today, OpenAI faces a “scale paradox”: usage growth directly inflates inference spend faster than revenue in some segments, while frequent model upgrades reset training costs. At larger scale (e.g., $100B+ revenue), fixed infrastructure efficiencies, caching/provisioning, and cheaper inference hardware could lift gross margins toward 50%+, reducing relative burn—yet the need for frontier leadership keeps capex structurally high (projected hundreds of billions cumulatively). Competitors or open-source alternatives could compress pricing further, pressuring the economics.[21]

This dynamic favors players with diversified infrastructure, strong enterprise lock-in, or differentiated efficiency; new entrants would need substantial capital or novel cost advantages to compete on the same frontier.

---

Recent Findings Supplement (July 2026)

OpenAI’s Q1 2026 cash burn reached $3.7 billion on $5.7 billion in revenue (roughly 65% of revenue), per The Information’s June 2026 reporting on internal documents shared with shareholders.[1][2] This reflects continued heavy infrastructure scaling amid missed internal user and revenue targets, raising internal concerns from CFO Sarah Friar about funding future data-center contracts.[3]

• 2025 audited financials (leaked June 2026 via Ed Zitron and verified by Financial Times): Revenue $13.07 billion; cost of revenue $7.5 billion; R&D $19.18 billion (including ~$10.59 billion paid to Microsoft, likely for training/compute); sales & marketing $5.73 billion; G&A $1.57 billion. Total expenses ~$34 billion; operating loss $20.92 billion; net loss attributable to the company ~$38.5 billion after noncontrolling interest adjustments (impacted by for-profit conversion).[4][5]
• Sacra estimates (updated 2026): Annualized revenue hit $25 billion by February 2026 (from $20 billion end-2025); 2025 gross margin 33%; inference costs $8.4 billion (2025), projected $14.1 billion (2026); cash burn projected at ~$27 billion (2026) and ~$63 billion (2027). Not expected to turn cash-flow positive until 2030.[6]
• Internal projections cited in mid-2026 reporting: ~$14 billion net loss for 2026; no profitability until 2029–2030.[7]

This burn rate and loss trajectory underscore extreme capital intensity driven by inference and training compute, with Microsoft Azure as the primary (and increasingly dominant) provider. OpenAI accounts for ~45% of Microsoft’s $625 billion cloud backlog and has committed up to $250 billion in Azure spend over time.[8] Efficiency gains (software optimizations halving inference costs on some models) are emerging but have not yet offset overall spend growth.

Gross margins deteriorated in 2025 due to quadrupled inference costs but show segment differentiation and recent recovery. Adjusted gross margin (revenue minus inference) fell to 33% in 2025 from 40% in 2024, missing an internal 46% target; free-user traffic significantly drags overall figures.[9][10]

• Compute/inference margin (share of revenue after server costs) improved to ~70% by October 2025 on paying segments (from ~35% in early 2024) via optimizations; company-wide figures are lower due to free users.[11]
• By end-Q1 2026, adjusted gross margin recovered to ~39%; targeting 52% by year-end 2026 through continued software-driven cost cuts.[12]
• Segment differences: API and enterprise deals generally support higher margins than consumer ChatGPT subscriptions (especially free tier or heavy Pro usage, where losses have been acknowledged on premium plans). Enterprise now >40% of revenue mix.[6][13]
• Inference cost per query trends: No granular per-query dollar figures in newest reports, but efficiency improvements (e.g., >50% cost reduction on select models via software alone) are trending downward even as absolute spend rises with volume.[12]

Headcount and compensation costs remain elevated due to talent competition and high valuations. OpenAI ended 2025 with ~7,850 employees (targeting 8,000 by end-2026) and provides average equity packages of ~$1.5 million per employee—the highest recorded at a private tech startup.[14][15]

• Compensation is a material but secondary driver compared to compute; R&D (including equity) dominates expenses.
• Growth in headcount supports enterprise push and product development but adds to operating burn.

Capital intensity remains high in the current model but has pathways for improvement at scale through efficiency and mix shift. Current operations require massive, front-loaded Azure commitments and cluster buildouts that scale with usage and model size, leading to inference often exceeding 50% of revenue in recent periods and overall losses far outpacing revenue.[16]

• At scale: Software optimizations (already halving costs on models), better hardware utilization (e.g., newer NVIDIA platforms), higher enterprise/API mix (>40% and growing, with potentially superior unit economics), and pricing power could lift gross margins toward or beyond the 52% 2026 target. However, training frontier models and serving multimodal/long-context workloads will sustain high absolute capex intensity.
• Competitor/entrant implication: Pure inference plays face similar economics unless they secure differentiated compute access or focus on narrow, high-margin verticals; the data/compute moat (real-time usage insights for optimization and underwriting-like advantages in related products) is hard to replicate without equivalent scale.

Cost Structure Breakdown (2025 Actuals, approximate; sources: leaked financials and Sacra estimates)

Category	Amount (USD)	Notes
Revenue	$13.07B	Core subs + API/enterprise (excludes some Microsoft licensing/one-offs per Sacra)
Cost of Revenue (incl. inference)	$7.5B	Inference ~$8.4B cited separately in some estimates; Microsoft payments included
R&D (incl. training/compute)	$19.18B	Heaviest line; ~$10.59B to Microsoft for compute/R&D
Sales & Marketing	$5.73B	Scaling enterprise push
G&A	$1.57B	Overhead
Total Expenses	~$34B	Operating loss ~$20.92B before adjustments
Net Loss (attributable)	~$38.5B	After noncontrolling interest and conversion impacts

Implications for competitors/entrants: The economics favor those with access to subsidized or efficient compute (e.g., via hyperscaler partnerships) or narrower scopes that avoid broad free-tier drag. Sustained inference cost reductions and enterprise mix shifts are critical for any path to breakeven; otherwise, burn rates will continue pressuring valuations and funding needs. Data as of mid-2026 shows improving efficiency signals but no fundamental shift in capital intensity yet.