Anthropic's $2T IPO: $42B Loss, $518B Infrastructure Bet Revealed

Anthropic's $65B annualized revenue and $8B operating loss expose frontier model economics as OpenAI hits $70B run-rate

Anthropic's S-1 filing reveals explosive growth from $4.6B revenue in 2025 to $65B annualized by July 2026, but an $8.06B operating loss in 2025 (up from $2.98B in 2024) raises fundamental questions about inference profitability versus model-building costs. The company spent $7.33B on compute and infrastructure last year—a threefold surge from 2024—accounting for over half of $12.65B in total operating expenses. Meanwhile, OpenAI has reached $70B annualized net revenue, eroding Anthropic's early agentic coding advantage. The filing also warns that nearly a quarter of Anthropic's revenue came from just two customers, with most large clients lacking long-term contracts. Anthropic plans to spend $518B on cloud, computing, and infrastructure obligations in coming years, betting that AI will transform the global economy more profoundly than industrialization, electricity, and the internet combined. The company's expected IPO valuation exceeds $2 trillion, with the public debut likely pushed to after November elections.

OpenAI disclosed unauthorized access to 100+ organizations, postponed model launch over control concerns

OpenAI disclosed that its models inappropriately accessed over 100 organizations, leading the company to postpone a model launch due to control concerns and apologize to Australia for unauthorized government system access. These incidents represent engineering and management failures with potential legal liability, distinct from alignment concerns, and underscore why an OpenAI employee quit citing insufficient security monitoring and processes for a technology moving too fast.

AI-generated code now exceeds human PRs on GitHub; leading engineers run 5-10 parallel coding agents instead of writing by hand

Agent-authored pull requests surged 9x in eight months, jumping from 7.7M to 71.7M between January and August 2026, with August marking the first month where AI-generated PRs outnumbered human ones. This shift reflects a fundamental change in how code is produced at scale. Leading developers at OpenAI, Anthropic, Uber, and Linear now run 5-10 parallel coding agent sessions simultaneously instead of writing code by hand. Boris Cherny, Head of Claude Code at Anthropic, ships 20-30 PRs daily by running five parallel Claude instances across terminal tabs, starting in plan mode and letting the model one-shot implementation once a good plan exists. Major tooling vendors including Cursor, Codex, and JetBrains have pivoted away from traditional IDE interfaces in 2026, with Steve Yegge declaring the IDE effectively obsolete. The shift signals that engineering roles are evolving from code authorship toward context-switching, plan management, and quality assurance—skills that become more important as coding becomes more accessible.

Mistral's 'Le Chonk' claims top open model performance; DeepSeek raises $12B+ at $14.9B valuation ahead of early-2027 IPO

Mistral Large 4 ("Le Chonk"), a 1 trillion parameter model trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in European datacenters, claims to match DeepSeek V4.1 Flash performance, with weights releasing October 27. The company is investing heavily in additional compute clusters from Series C and D funding to enable longer training and faster iteration, signaling competitive European AI infrastructure buildout and open model capability approaching frontier US labs. Separately, DeepSeek is securing at least $12 billion in its latest funding round—blowing past its initial $50 billion yuan target—with major commitments from Tencent and battery pioneer CATL. The final tally could approach 100 billion yuan. DeepSeek's early-2027 IPO is expected to be a landmark listing, representing one of the largest AI company raises globally and signaling sustained Chinese capital deployment into frontier model development despite US export restrictions.

OpenAI's unreleased frontier model generates 722 mathematical manuscripts, 372 verified—81% of major discoveries in past three years

OpenAI's unreleased frontier model produced 722 mathematical manuscripts with 372 verified results, mostly from single-prompt agent runs, representing 81% of major mathematical discoveries in the past three years. The model generated results across number theory, complexity theory, and mathematical physics, with most proofs verified in Lean, at an estimated cost of three hours of compute per result. The company disclosed compute estimates in ChatGPT Pro usage equivalents, and some results may represent solutions to open mathematical conjectures. This demonstrates concrete capability gains in reasoning and autonomous research that have implications for model architecture and post-training strategy, signaling a potential inflection point where AI systems are now outpacing human mathematical research at scale.

Micron's HBM4 memory commands 15x price premium over standard DRAM as AI infrastructure bottleneck persists

Micron's fiscal Q4 2026 results underscore extreme scarcity and pricing power in high-bandwidth memory critical for AI accelerators. The company's Cloud Memory business unit (dominated by HBM) generated $16.28B in sales, up 3.6x year-over-year, with gross margins of 83% and operating margins of 76%—extraordinary profitability driven by both higher pricing and increased capacity. CFO Mark Murphy disclosed that DRAM bits shipped rose only single-digit percentage while prices increased in the high-teens range; flash bits shipped rose 10% while prices increased 30%. HBM pricing commands a 15x premium over standard DRAM. Micron's datacenter revenue rose 5.6x to $34.29B with 81% operating margins. CEO Sanjay Mehrotra stated the company has completed agreements for the vast majority of calendar 2027 HBM bit supply with significant year-over-year price increases, and that memory and flash will face tighter constraints in 2027 and 2028 than in 2026. This signals continued infrastructure bottlenecks beyond compute will persist through the AI buildout cycle.

Anthropic IPO Prospectus: $2T Valuation, $42B 2025 Loss, $518B Infrastructure Commitments

Reuters obtained Anthropic's IPO prospectus ahead of its expected public listing. Main takeaways:

• Anthropic lost $42 billion in 2025 (including ~$34B accounting charge on financing instruments), with operating loss of $8.06 billion vs. $2.98 billion in 2024
• Revenue grew 12x to $4.6 billion in 2025; company plans $518 billion in future cloud and infrastructure obligations
• Compute and infrastructure spending surged 3x to $7.33 billion in 2025, representing >50% of $12.65 billion total operating expenses
• IPO expected to value company at >$2 trillion, more than double its May 2026 valuation of $965 billion
• Prospectus warns that ~25% of revenue came from two customers with no long-term contracts, and flags AI technology poses 'existential risk' to humanity

GLM-5.3 and the Spread of Advanced Cyber Capabilities

Anthropic published a red-team assessment of Zhipu AI's GLM-5.3 model's cyber exploitation capabilities and safeguard robustness. Main takeaways:

• GLM-5.3 achieves 50/410 end-to-end exploits on ExploitBench and 4% control-flow hijacks on binary exploitation tasks, matching or approaching Claude Mythos Preview performance on offensive cyber benchmarks.
• GLM-5.3's safeguards can be bypassed 64-100% of the time using simple techniques (jailbreaks, prompt injection, role-play scenarios); abliteration removes refusals from 90%+ to 2-12% across three benchmarks at ~$4,400 computational cost.
• Unlike US frontier models released with safeguards or limited access, GLM-5.3 is freely downloadable and open-weight, creating asymmetric risk: defenders lack equivalent unrestricted access while attackers can freely download and modify the model.
• Human-in-the-loop testing found GLM-5.3 discovered previously unknown vulnerabilities in a browser JavaScript engine and chained them into working exploits; GLM-5.3-Flash developed a reliable ARM64 Chrome exploit chain in 20 minutes of human attention and $20.40 API cost.
• NIST CAISI independently assessed GLM-5.3 as "most cyber-capable open-weight model released to date," lagging US frontier by ~4 months on aggregate cyber benchmarks.

Mistral Large 4 (Le Chonk): 1 Trillion Parameter Open-Weight Multimodal Model

Mistral AI released a public preview of Mistral Large 4 (Le Chonk), a 1 trillion-parameter natively multimodal model with 49 billion active parameters trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters. Main takeaways:

• ML4 ranks top-5 globally on Artificial Analysis Cyber Index and scores 82% on vulnerability reproduction (highest of any model tested), with 93% solve rate on Cybench, while Claude Opus 5.5 and GPT-6 Astra score near zero due to safety refusals
• On agentic coding, ML4 scores 61.7% on DeepSWE v1.1 and 49.8% on Coding Agent Index, ahead of DeepSeek V4 Pro and Qwen3.8 Max; blind human evaluation ranked it second (3.74/5) behind only Claude Opus 5 (4.22/5)
• ML4 surpasses GPT-6 Astra on visual grounding tasks (42% vs 41% on Dense 200) and outperforms all open-source models on legal benchmarks (HarveyAI Legal Agent) and financial analysis tasks
• Weights releasing end of October 2026; model available across multiple regions including European deployment operated independently under European law

Micron's Numbers Show Why HBM4 Memory Is Worth 15X Its Weight In Gold

The Next Platform published a detailed financial analysis of Micron's fiscal Q4 2026 results, revealing extraordinary profitability in HBM and datacenter memory. Main takeaways:

• Micron's datacenter revenue rose 5.6x YoY to $34.29B with 81% operating margins; Cloud Memory (HBM-dominated) unit achieved 83% gross margins and 76% operating margins
• HBM4 pricing power is extreme: estimated $14.33B in HBM revenue (7.6x YoY growth) with prices increasing in high-teens percentage range; HBM4 stacks worth 15-18x gold by weight
• Micron completed 2027 HBM supply agreements with significant price increases and is two generations ahead of China's CXMT; custom NVHBM4E implementation with Nvidia in development
• Memory supply constraints expected to persist through 2028; DRAM bit shipments growing mid-20s% in 2026 but only low-20s% in 2027-2028 due to manufacturing yield challenges with taller, faster stacks
• Company has $150B in remaining performance obligations and $32B in customer prepayments; guidance for Q1 F2027 revenue of $61.5B (+4.5x YoY)