2025–2035 AI Industry Profitability & Collapse Forecast Report

Executive Summary

Bottom line. Frontier‑model AI can become sustainably profitable—but only if unit economics improve at least as fast as usage grows. The decisive levers are: (i) lowering compute cost per useful token/inference (custom silicon, higher utilization, better model design and caching), (ii) locking in cheap/firm power, (iii) shifting mix toward high‑margin enterprise and platform revenues, and (iv) stretching training amortization (less frequent “big‑bang” pre‑trains, more continual/mixture‑of‑experts upgrades).

Projected first year of GAAP profitability (base cases):

Anthropic: FY2027 (management target referenced in reporting).
OpenAI: FY2029 (widely reported base case); upside to FY2027–28 if cost curves and enterprise monetization break right.
Google DeepMind (Alphabet): Alphabet is already profitable; AI is accretive but capex heavy. Profitability is managed at Alphabet/Google Cloud level, not as a standalone P&L.
xAI: FY2029 base; upside FY2028 with strong subscriber/API and ad tie‑ins; otherwise slips beyond 2031. (Funding substantial, revenue still developing.)

If profitability is delayed: Most exposed are independent labs with heavy compute commitments and thin enterprise mix. With no further financing, OpenAI’s mid‑2025 cash of ~$17.5B and 2025 burn target of $8.5B imply ~2 years of runway (to ~mid‑2027). The NVIDIA–OpenAI 10 GW / up to $100B strategic deal (LOI) would materially extend runway and tie capex to staged deployment milestones.

Systemic contraction (stress scenario 2026–2028): Would dent Azure/AWS/Google Cloud growth, stall AI‑led productivity tools, and soften semiconductor capex—while easing near‑term grid stress. It would also strengthen state‑backed or hyperscaler‑embedded labs (Alphabet, Microsoft‑OpenAI, AWS‑Anthropic) and intensify geopolitical competition, especially as China’s low‑cost models pressure Western pricing.

What must happen to avoid that outcome:

Target unit economics: drive provider cost for GPT‑4o/Claude‑class capability toward ≤$0.05 per 1M tokens all‑in by 2027 on mainstream queries, with API gross margin ≥45% and inference GPU utilization ≥60%.
Lock power and cooling: secure long‑dated PPAs at $40–$50/MWh equivalent, PUE ≤1.15, and build closer to low‑cost generation.
Shift revenue mix: grow enterprise seats/platform embeds to >50% of revenue; expand vertical agents (coding, customer service, finance) with >120–130% net revenue retention.
Cadence discipline: major pretraining every 12–18 months, with MoE and reasoning‑sparring to flatten compute growth.

Download the data tables and scenario files used in this brief:

Financial snapshot: CSV
Power & infrastructure pipeline (indicative): CSV
Profitability scenarios (illustrative): CSV
Runway sketch (very rough): CSV

A quick visual of illustrative profitability windows (optimistic → pessimistic) is embedded above.

1) Current Financial Overview

OpenAI.

Revenue: ~$4.3B in H1‑2025; FY2025 target: $13B.

Costs/Losses: H1‑2025 R&D ~$6.7B; cash burn ~$2.5B in H1; full‑year burn target $8.5B; cash & securities ~$17.5B at H1‑end.

Profitability outlook: FT reporting indicates not profitable until ~2029 under then‑assumptions.

Infrastructure posture: pursuing multi‑site mega‑DC buildout (e.g., Abilene, TX gas plant ~360.5 MW for a first “Stargate” site; UAE site exploration).

Strategic finance: NVIDIA LOI to invest up to $100B and deploy ≥10 GW of systems from 2026.

Anthropic.

Revenue: >$5B ARR by Aug‑2025 (company).

Burn / break‑even: 2024 cash burn ~$5.6B; 2025e ~$3B; targets cash‑flow positive by 2027 (per reporting on management plans).

Funding: $13B Series F at $183B post‑money (Sept‑2025) plus Amazon’s $4B (2023–2024) strategic investment; primary cloud is AWS/Bedrock.

Google DeepMind (Alphabet).

Alphabet’s FY2024 R&D ~$49B; 2025 capex guidance now $91–$93B (DCs, TPUs, power). Google Cloud revenue growth and $155B backlog signal AI‑driven demand. (DeepMind results are not broken out.)

xAI.

Funding: $10B at ~$200B valuation (Sept‑2025).

Infrastructure: Memphis plan for ~422 MW gas turbines to power DCs; broader “Colossus” plans are owner‑stated and subject to approvals.

Revenue: press/analyst estimates vary and remain speculative at this stage.

Note on pricing (for revenue quality): The direction of model API pricing continues to fall (e.g., OpenAI pricing page, Anthropic Sonnet 3.5/4.5 at ~$3 per 1M input / $15 per 1M output, Gemini Flash‑Lite double‑digit cents per 1M tokens), intensifying the need to cut provider costs at least as fast.

2) Power and Compute Cost Analysis

Macro context.

U.S. wholesale power is projected around $40/MWh (2025 average), with regional spikes; U.S. electricity demand to hit records in 2025–2026, driven by AI data centers. Capacity prices in PJM have spiked nearly 10× vs. 2024/25 in later auctions—grid constraints are binding.

HBM and advanced packaging remain tight (HBM supply sold out into 2025; CoWoS capacity expanding but still gating near‑term ramps).

Scale indicators (selected, mainly planned capacity):

OpenAI–NVIDIA partnership: ≥10 GW of NVIDIA systems to be deployed from 2026; at 90% capacity factor, that equates to ~78,840 GWh/year of electric load and $3.15–$5.52B/yr in power cost at $40–$70/MWh, before T&D and fuel hedges.
OpenAI Abilene (first “Stargate” site): ~360.5 MW gas plant; ~2,842 GWh/year at 90% CF ⇒ ~$110–$200M/yr power cost at $40–$70/MWh.
xAI Memphis turbines: ~422 MW ⇒ ~3,327 GWh/year at 90% CF ⇒ ~$130–$230M/yr.
Alphabet/Google: signed ~8 GW of new clean energy contracts in 2024; company reports 12% drop in DC emissions in 2024 despite rising load—reflects aggressive PPA strategy; electricity use rose ~27%.

Inference cost pressure. Pricing is falling faster than most providers’ costs absent architectural gains. Providers must rely on: (i) MoE (route only a fraction of parameters per token), (ii) caching and KV‑reuse, (iii) quantization/compilers, and (iv) custom silicon to keep provider‑side cost per 1M tokens comfortably below market prices as Gemini/DeepSeek/others drive down price points. Google’s addition of very cheap “Flash/Flash‑Lite” models illustrates this pressure.

3) Profitability Roadmap (scenarios)

Key shared levers across companies

Chip & systems cost curves: Transition to B200/GB200 and successors; secure HBM3E→HBM4 supply; lift utilization via schedulers/agents; expand CoWoS packaging supply.
Energy price and availability: Lock PPAs around $40–$50/MWh; site near firm generation (NG, hydro, nuclear pilots).
Revenue mix & pricing: More enterprise seats, platform rev‑share, and vertical agents; keep usage margin positive as list prices compress.
Model cadence: Fewer full retrains; more continual training, MoE, reasoning tokens only when necessary.

Scenario estimates (summary) — details in the attached CSV:

Company	Optimistic	Base Case	Pessimistic	Breakeven mechanics
OpenAI	FY2027–28	FY2029	FY2031+	10–20× cost/throughput improvement vs. GPT‑4o baseline, enterprise attach, lower Azure revenue‑share, power PPAs
Anthropic	FY2026–27	FY2027	FY2029+	AWS Bedrock distribution, Claude Code/agents seats, Trainium2 efficiencies
DeepMind/Alphabet	(Alphabet profitable)	n/a	n/a	Maintain Cloud margins; monetize Gemini/Vertex while funding TPUs
xAI	FY2028	FY2029	FY2031+	Scale X subscriber/API revenue, ads, staged DC build at target $/token

(Our visual “profitability windows” plot is shown above.)

4) Failure and Collapse Scenarios

Runway calculations (very rough): see downloadable table.

OpenAI: $17.5B cash/securities at H1‑2025 vs $8.5B burn guide ⇒ ~2 years runway to ~mid‑2027 if no new financing and costs scale with plan. The NVIDIA LOI (up to $100B, staged vs. deployed GW) would significantly reduce financing risk and may re‑shape unit costs (supply assurance, possible volume economics).

Anthropic: $13B Series F plus strategic capital (Amazon) and growing ARR create a multi‑year runway; management target break‑even by 2027.

DeepMind (Alphabet): Not an independent failure candidate; risks are margin compression and capex burden at Alphabet as capex lifted to $91–$93B.

xAI: $10B raise funds build‑out; project power ~422 MW; revenue still early. If monetization lags and no follow‑on funding, 2–4 years of runway depending on capex timing.

What “failure” looks like in practice.

Balance‑sheet or cash‑flow failure: default on DC and chip purchase commitments; loss of cloud credits; emergency down‑rounds.
Orderly restructuring: asset sale to hyperscaler; IP licenses; pivot to open‑sourcing to stem costs and keep ecosystem goodwill.
Receivership of projects vs. company: sovereign buyers (national labs) may acquire partially built AI factories; customers migrate to hyperscaler‑bundled models.

Shock channels to model:

Investor pull‑back: Kills speculative ramps; lowers GPU orders (esp. Blackwell), cooling HBM pricing; capex at clouds re‑prioritized.
Power shocks: PJM/CAISO congestion or fuel spikes; local moratoria; capacity charges up materially.
Export controls / supply chain: HBM/CoWoS bottlenecks or policy restrictions delay ramps; vendors rebalance allocations.

5) Global Economic & Societal Impact if AI Companies Fail

Cloud providers: AI contributes meaningfully to Azure/Google Cloud growth; Alphabet just raised 2025 capex to $91–$93B citing AI demand—an AI downturn would remove a major growth driver, tighten margins, and defer DC build.

Jobs & productivity: Delayed AI adoption slows roll‑out of coding assistants and back‑office automation, dampening projected productivity gains; tool vendors reliant on API credits would consolidate.

Semiconductors & supply chains: HBM and CoWoS suppliers face whiplash; however, a pause eases grid/power pressure and may normalize memory prices.

Geopolitics: China’s low‑cost LLMs and aggressive pricing (e.g., DeepSeek/01.ai) put continued pressure on Western pricing; a U.S. pullback would encourage state‑supported and hyperscaler‑embedded champions. Western governments would likely expand procurement and loan guarantees to sustain strategic AI capacity.

Energy & environment: An industry slowdown would reduce near‑term load growth (IEA/EIA outlooks), but stranded or delayed energy assets (gas turbines, PPAs) could produce cost overhangs for utilities.

6) Strategic Recommendations (with measurable indicators)

A. Unit economics & pricing

Set 2027 targets: provider cost for GPT‑4o/Claude‑class inference ≤$0.05 per 1M tokens on mainstream tasks; API gross margin ≥45%; enterprise margin ≥60%.
Adopt hybrid pricing: tiered per‑token + seat‑based enterprise SKUs; bundle agents and domain adapters rather than pure tokens; success metric: enterprise ARR mix >50%.

B. Compute efficiency

MoE everywhere: keep active params/token low; push KV‑cache reuse across sessions; target GPU‑hour per 1M tokens ↓ by 5–10× over 24 months (roadmap metric).
Compiler/quantization stack: standardize INT4/FP8 inference where quality allows; KPI: serving throughput +3–5× on same silicon by 2027.

C. Silicon & supply

Dual‑source chips: hedge NVIDIA with custom chips (Broadcom/TSMC for OpenAI; Trainium2 for Anthropic; TPUs for Google); measure: ≥30% of inference on non‑H100/B200 by 2027 without quality loss
HBM/CoWoS contracts: secure multi‑year allocations; KPI: no >10% shortfall vs. plan during ramps

D. Power & siting

Lock PPAs at $40–$50/MWh with firm capacity; site near generation; PUE ≤1.15; renewables share ≥80% by 2028; pilot SMRs/geothermal where feasible (Alphabet initiative referenced)
Grid partnership playbook: fund local capacity upgrades to avoid PJM‑style capacity spikes; KPI: $/kW capacity charges stable within ±15% year over year

E. Product & cadence

Training cadence discipline: major pre‑train every 12–18 months; more continual learning and reasoning‑on‑demand to cap compute growth
Platform moats: secure OEM embeds (office suites, IDEs, CRM, ERP); KPI: NRR ≥130%, paid seats CAGR >40%

F. Risk & governance

Counterparty diversification: avoid single‑cloud lock‑in on DC/credit; KPI: no single partner >50% of compute availability
Regulatory posture: pre‑wire audits/safety reporting to reduce compliance friction; KPI: <1% revenue at risk from regulatory interruptions

Leading indicators of "approaching sustainability": Provider cost/1M tokens (quality‑adjusted) trending toward $0.01–$0.05; GPU utilization ≥60%; API gross margin ≥45%, enterprise margin ≥60%; capex/revenue <1.2× by 2029; cash conversion positive; model refresh interval ≥12 months on average.

7) Company‑by‑Company Profitability Outlook

OpenAI

Base case: FY2029 profitability, assuming cost/throughput keeps pace with falling prices and enterprise mix expands. Upside: FY2027–28 if the NVIDIA 10 GW program and custom silicon materially reduce unit cost and if long‑dated PPAs land near $40–$50/MWh. Downside: FY2031+ if energy, HBM, or policy delays persist; cash runway without new funds ~mid‑2027; with staged NVIDIA financing, runway extends alongside buildout.

Anthropic

Base/management case: FY2027 break‑even as ARR scales on Bedrock/Vertex distribution and Claude Code/agents. Upside: FY2026–27 with Trainium2 and enterprise seat growth. Downside: FY2029+ if inference prices reset lower or AWS silicon cost curves slip.

Google DeepMind (Alphabet)

Alphabet‑managed profitability; 2025 capex $91–$93B indicates continued AI infra expansion. Risk is margin compression rather than survival; strong Cloud backlog supports demand.

xAI

Base: FY2029; Upside: FY2028 with high ARPU on X and enterprise APIs; Downside: FY2031+ if subscriber and ad monetization lag and Memphis power comes online late.

Appendices & Tables

The following files contain the structured tables used in this analysis:

Financial snapshot (FY2025 YTD & notable items): Download
Power & infrastructure pipeline (indicative): Download
Profitability scenarios (per company): Download
Runway sketch (illustrative only): Download
Power‑to‑cost translation for select plans (OpenAI/xAI): Download

Important notes on uncertainty. Many frontier‑lab figures are private; where we used press/management disclosures we label them; where we extrapolated, we state assumptions (e.g., 90% capacity factor for back‑of‑the‑envelope power costs).

Sources (selected citations)

OpenAI H1‑2025 financials & 2025 targets: Reuters summary of shareholder disclosures (H1 revenue ~$4.3B; burn target $8.5B; R&D $6.7B; cash $17.5B)
OpenAI profitability not before ~2029: Financial Times
OpenAI data centers / Abilene 360.5 MW / UAE site: Business Insider; Reuters (site strategy)
NVIDIA–OpenAI strategic deal (≥10 GW; up to $100B): Reuters; NVIDIA press release
Anthropic burn/break‑even 2027; run‑rate revenue: Reuters; Anthropic newsroom; The Verge coverage of Series F
Amazon–Anthropic partnership (Bedrock): Amazon announcement
Alphabet capex/backlog & Cloud growth: Reuters (capex $91–$93B; backlog growth)
Google environmental/energy reporting: 2025 Environmental Report—PPAs and emissions trends; coverage summarizing 12% emissions drop with rising load
xAI funding and Memphis turbines (422 MW): Reuters
HBM/packaging constraints: Reuters/TrendForce on HBM sell‑outs; TrendForce on CoWoS expansion
Power & grid pricing context: EIA wholesale price outlook; EIA load records; IEEFA on PJM capacity spikes
Pricing pressure from low‑cost models: Reuters on Google's "Flash/Flash‑Lite" to counter low‑cost players