Anthropic quietly filed its S-1 with the SEC on June 1, 2026. Four days later, the rumored valuation circulating through private market channels hit $965 billion. The algorithm priced the ape before the crowd did. Liquidity didn't wait for the official announcement.
That headline number—$965 billion—is not the story. The story is the structural divergence beneath it. It's the difference between OpenAI's $852 billion infrastructure-heavy bet, Anthropic's leveraged financial engineering, and the Chinese labs calibrating their exits to sovereign capital cycles. This is not one AI IPO wave. It is three distinct capital strategies, and only one of them survives the next bear market intact.
Context: The Timeline Tells You Everything
The 2026 IPO calendar is alphabet soup of delayed S-1 filings and confidential submissions. OpenAI reportedly filed with Goldman Sachs and Morgan Stanley, sliding its public debut to 2027. Moonshot AI, riding the momentum of Kimi K3's $300 million ARR, is reportedly targeting a $50 billion valuation in Hong Kong. DeepSeek, backed by Tencent and CATL, is rumored to be moving toward a STAR Market listing.
Why now? Because model capability differentiation is commoditizing faster than most investors want to admit. The GPT-4 to GPT-4o to o1/o3 trajectory shows a lead window that has collapsed from 12-18 months to roughly 6-9 months. DeepSeek-R1 demonstrated that frontier-adjacent performance can be achieved at a fraction of the cost. When the technical gap narrows this quickly, the marketable difference becomes structural. Capital density replaces architectural genius.
The core question isn't who trains the best model. It's who can lock in the cheapest inputs—compute, energy, capital, and regulatory permission—for the longest duration.
Core: The Hidden Battlefield Is the Balance Sheet
OpenAI's strategy is a $105 billion infrastructure commitment in Ohio. Let me be precise about what that means. Based on my audit experience, this is roughly 4.25 gigawatts of power capacity with 20-year land leases. It's a bet that model performance is less durable than compute access. The market reads this as 'vertical integration' but the financial structure says something different: an attempt to convert volatile future compute costs into fixed, depreciable assets.
The math doesn't sit comfortably. At an $852 billion post-money valuation, assuming a hypothetical $25 billion in annualized revenue, OpenAI trades at ~34x static price-to-sales. That's a multiple that demands a 60%+ growth rate sustained for years. The infrastructure build-out, however, has a deprecation cycle of decades. The unit economics of AI training clusters do not improve on a linear curve. They hit power constraints, transformer delivery bottlenecks, and cooling system physics. 4.25GW of power capacity is a massive commitment but grid interconnection timelines routinely exceed data center construction schedules.
Anthropic's approach is what I call 'clean balance sheet, dirty structure.' They're reportedly using a $71 billion off-balance-sheet debt vehicle managed by Apollo and Blackstone. Directionally, this means their core financial statements will show healthier margins and lower leverage than reality. The 9650亿美元的 headline is pure noise if you don't understand that the loan structure has priority claims over equity. The real question is what the common equity is worth after the debt holders get paid.
Their P/S math is worse. If you assume no more than $20 billion in revenue and a $965 billion valuation, the static ratio sits between 45x and 50x. That implies the market is pricing zero risk of disruption for the next three years. When I ran my algorithms across this data, I saw a conflict already brewing. The technical benchmark gap on SWE-bench—where Claude reportedly leads by 8-10 points—simply doesn't justify this massive a discount to every other financial metric in the sector.
The Moonshot and DeepSeek Exceptions
China's AI labs are playing a different game. Their accounting bears little resemblance to the American model. Kimi K3's $300 million ARR against a $50 billion valuation target creates a P/S ratio exceeding 166x. In any mature market, that number is a death sentence. But on the Hong Kong exchange, scarcity has its own discount rate. The Hong Kong market gives a premium for being the 'first pure-play Chinese foundation model.' That is not a valuation. It is a reward for being in the right place at the right time.
DeepSeek's $74 billion valuation and their backing by Tencent and CATL suggest a different protection mechanism. Despite the export controls that severely limit access to advanced Nvidia chips, the sovereign-aligned capital structure provides something no American lab has: patience combined with state-adjacent procurement.
But this dual-track structure—US labs pursuing infrastructure and financial engineering, Chinese labs pursuing domestic ecosystem monetization—creates a fundamental investor blind spot. Everyone is comparing pricing power and model grades. No one is analyzing data provenance.
Contrarian: The Copyright and Compliance Time Bomb No One Is Pricing
The dominant narrative on Anthropic's $1.5 billion copyright settlement was that it cleared the path to IPO. Structure is not a cage; it is a launchpad. I would argue the opposite. The settlement signals that content provenance has become a non-negotiable input, not a legal afterthought. It's a fixed cost that sets a precedent.
OpenAI is still mired in copyright litigation across multiple jurisdictions, most notably the New York Times case. If you scrutinize the risks in Anthropic's F-1, you will find a solvency risk that no one is talking about: what happens when the next training run requires a new data source and triggers a new settlement cycle? The EU AI Act's General-Purpose AI obligations are now fully active. The compliance burden is less a 'cost of doing business' and more a 'structural tax' that fundamentally favors incumbents. Small startups now face fixed compliance costs that genuinely threaten their ability to operate.
This is where my Celsius collapse pre-mortem framework applies. In mid-2022, I flagged a 15% discrepancy in reserves using a standardized audit framework. Right now, the AI sector has a similar mismatch. The difference is between reported 'technical capability' and the unaudited 'legal and regulatory liquidity' to maintain that capability. Anthropic's decision to abandon the Decart acquisition tells me that even AI labs flush with cash know their limits when it comes to technical integration and due diligence. That is a red flag for future M&A-led growth narratives.
The Real Risk: An Unpriced Down Round
The public narratives describe OpenAI as 'infrastructure safety,' Anthropic as 'financial engineering efficiency,' and Chinese labs as 'sovereignty-aligned stability.' But the market is treating these private valuations as if they are already public prices. They are not. Value is a consensus, not a contract.
Here's the number no one is discussing: if Anthropic prices its IPO at the low end of the reported range and the stock trades below issuance in the first 90 days, it triggers a cascading repricing for the entire AI sector. Considering that Moonshot AI and DeepSeek have not yet gone public, they would face significant downward pressure on their offering prices.
Historical precedent supports this. Snowflake had an extreme P/S ratio in 2020, and it took four years for the markets to fully absorb that valuation. That was a software company with good gross margins. AI labs have hardware depreciation, energy costs, and unpredictable headcount expenses.
The tech industry quietly expects 'compensation shock' once labor costs are diluted and employee stock options become listed shares.
Takeaway: The Next Watch
First, follow the S-1 filings on the SEC website, not press leaks. The risk factors section in Anthropic's S-1 will be the single most valuable document published this year.
Second, watch for the first public-market down round. If a valuation cut happens, it will not be a minor event—it will be the signal that the capital allocation phase of AI is over and the Darwinian phase of cost discipline has begun.
Will the next AI bull market be won by the biggest model, or by the lab that can survive a capital winter? Liquidity is a ghost. Watch the volume.
