OpenAI's $38.5B Loss, Anthropic's $47B ARR, and GLM-5.2 Goes Open — AI Midweek Roundup

June 17, 2026 — This week has been one for the history books. Three stories broke within 72 hours that fundamentally shift how we understand the AI landscape. Let's dig into each one.

---

1. OpenAI's Audited Financials: $38.5 Billion Loss — and That's Not the Whole Story

On June 15, tech journalist Ed Zitron published audited financial documents for OpenAI, independently confirmed by the Financial Times. The numbers are staggering.

The top-line figures:

• 2025 Revenue: $13.07 billion (up from $3.7B in 2024 — a 3.5x increase)
• Total costs: $34 billion
• Operating loss: $20.92 billion
• Net loss attributable to OpenAI: $38.53 billion
• Gross net loss (before minority interest allocations): $60.35 billion

The operating loss of $20.92B is the more telling number for operational health. It means OpenAI spent $1.60 for every dollar it earned in 2025. The silver lining: that's down from $2.37 per dollar in 2024. The ratio is improving as revenue scales.

Where the Money Goes

• R&D: $19.18 billion — the largest line item. Training ever-larger models isn't getting cheaper.
• Azure dependency: $17 billion went to Microsoft for cloud compute. OpenAI is effectively a massive Azure customer.
• Restructuring charge: $41.55 billion in non-cash charge from the for-profit conversion completed October 28, 2025. This is an accounting item, not actual cash — but it will appear in OpenAI's eventual S-1.

The Big Picture

OpenAI's revenue tripled but its costs nearly tripled too. The company confidentially filed its draft S-1 with the SEC just 8 days before the leak. For IPO investors, these numbers are a double-edged sword: incredible growth on one side, mind-boggling losses on the other. The bull case is that the loss-to-revenue ratio is improving. The bear case is that $17B in Azure spend alone shows how dependent OpenAI is on Microsoft's infrastructure — and how hard it would be to escape that dependency.

---

2. Anthropic Hits $47B ARR — and Surpasses OpenAI for the First Time

While OpenAI's financials dominated headlines, Anthropic quietly crossed a line that would have seemed impossible 18 months ago.

The growth curve:

• January 2025: $1B ARR
• April 2026: $30B ARR (surpassing OpenAI's ~$25B)
• May 2026: $47B ARR
• Growh: 47x in 16 months — the fastest in SaaS history

Series H: $65B raised at a $965B valuation, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital. For the first time, Anthropic's valuation ($965B) has surpassed OpenAI's ($852B after its March 2026 round).

Why Enterprise Is Betting on Claude

It's not about chatbots. Per the Ramp AI Index (June 2026) :

Claude Code is the rocket fuel. Key stats:

• 80% of Anthropic's own production code is now authored by Claude
• 4% of all public GitHub commits globally come from Claude Code — doubled in a single month
• 8 of the top 10 Fortune 10 companies are using Claude
• 1,000+ enterprise customers spending over $1M/year

Anthropic has confidentially filed for an IPO, likely targeting a public debut in late 2026. At $965B, that would be the largest tech IPO since Alibaba.

Consumer vs Enterprise: Two Different Worlds

In the consumer market, ChatGPT still dominates handily — 5.19 billion monthly visits vs Claude.ai's 952 million. But the enterprise market is completely inverted. The takeaway: consumer awareness doesn't equal enterprise revenue, and enterprise revenue is where the money is.

---

3. GLM-5.2 Goes Open: 753B Parameters, MIT License, Beats GPT-5.5

On June 13, Z.ai (智谱) dropped GLM-5.2 — and it's a genuine milestone for open-source AI.

The specs:

• 753B parameters (40B active — IndexShare MoE architecture)
• 1M token context (stable, not theoretical)
• MIT license — fully open, no regional restrictions
• 28.5T tokens training data
• Released on GitHub and Hugging Face

Benchmark performance:

• Terminal-Bench 2.1: 81.0 (vs GLM-5.1's 62.0 — a massive jump)
• Coding tasks: beats GPT-5.5 on FrontierSWE at 1/6th the cost
• Long-horizon tasks: competitive with Claude Opus 4.8

The IndexShare MoE innovation is the technical highlight: by reusing the same attention indexer across every four sparse layers, Z.ai cut per-token FLOPs by 2.9x at 1M context length. This is the kind of efficiency breakthrough that makes open-source models viable for production deployments.

GLM-5.2 is the strongest open-source model on standard coding benchmarks by a wide margin. Following MiniMax M3's release, it's another sign that Chinese AI labs are not just catching up — they're leading in specific dimensions.

---

What This All Means

Three stories, one week, pointing in different directions:

1. OpenAI is in a race against its own cost structure. $38.5B in losses, even with the non-cash charges stripped out, means the IPO window is critical. The company needs public market capital to keep training — there's no path to profitability without massive scale.

2. Anthropic has become the enterprise AI default. The 47x ARR growth in 16 months is unprecedented, but the real signal is the enterprise adoption depth. When 8 out of 10 Fortune 10 companies standardize on Claude, that's sticky revenue with high switching costs.

3. GLM-5.2 shows that open-source AI isn't slowing down. The gap between open and closed models continues to narrow. For developers and businesses building AI applications, the question is increasingly "why pay for proprietary?" rather than "can we afford proprietary?"

The takeaway: The AI landscape is not consolidating — it's fragmenting. OpenAI has the brand and consumer reach. Anthropic has the enterprise trust and growth curve. Open-source has the economics. All three are converging on the same reality: building AI infrastructure is brutally expensive, but the ROI for those who use it is becoming undeniable.

— Written for AI工具导航 by 未然