Anthropic, OpenAI's finances ahead of IPOs reveal computing cost challenges: report

OpenAI projects roughly $121B in spending for model training, driving massive cash burn and forecasting losses to continue into the 2030s. Anthropic faces escalating training costs too but records cloud-partner revenue and expects to break even sooner. Both firms are seeking large IPO raises to fund substantial cash needs, with bankers lobbying index rules to access bigger capital pools; Anthropic may IPO soon while OpenAI's timing is debated internally.

Analysis

The immediate, non-obvious consequence of outsized model-training costs is an acceleration of value capture toward hyperscalers and premium silicon providers rather than the LLM operators themselves. Hyperscalers can monetize both the compute dollar and distribution (reselling, embedding, and enterprise SLAs) which creates a two-sided moat: they earn gross margin on infrastructure while extracting sustainable take-rates from downstream model vendors.

A second-order supply-chain effect is rising bargaining power for custom-ASIC partners, power/cooling vendors, and data‑center real‑estate owners; these suppliers can effectively tax model builders through longer lead times and step-up pricing, making cheap scale a coordination problem not just a capital one. Conversely, productivity breakthroughs in algorithms (sparsity, retrieval-augmented methods, quantization) or a sudden influx of custom chips could compress opex per token by multiples and rapidly re-rate model operators’ cash flow outlook.

Timing matters: in the next 3–12 months capital raises and index inclusion discussions will be the dominant catalysts; over 12–36 months, model-efficiency gains or large enterprise contracts will determine who reaches self-funding. The asymmetric risk is a drawdown if open-source models and optimized stacks materially lower compute intensity — that would punish the incumbents priced for perpetual high demand while rewarding nimble, low-cost operators.

Key Decisions for Investors

• Long MSFT (buy Jan‑2028 LEAPS or add shares) — play hyperscaler pricing power and enterprise LLM contracts. Timeframe 12–24 months; objective return 50–100% if enterprise adoption accelerates, downside ~15–25% in macro drawdown (limited by enterprise sticky revenue).
• Long NVDA (buy shares + selective 12–18 month calls) — target exposure to continued GPU tightness and outsized pricing power. Timeframe 6–18 months; potential 2x+ upside if supply/demand stays tight, tail downside ~30% on inventory correction or competition from custom ASICs.
• Pair trade: Long AMZN (cloud exposure via Jan‑2028 calls) / Short PATH (UiPath) — prefer cloud infra over freemium/automation vendors that face margin pressure from compute costs and delayed monetization. Timeframe 12 months; expect AMZN to outperform PATH by 30–60% if enterprise deals scale, cut pair size to limit systemic risk.