The past few days have offered a vivid snapshot of where the AI industry actually stands heading into its next phase: enormous money flowing into frontier labs, hardware giants racing to own the "agent" layer, and Chinese open-weight developers shipping at a pace that keeps pressure on everyone else.

Anthropic's IPO filing reveals the scale of frontier AI economics

The most consequential disclosure this week came from Anthropic's draft S-1, reviewed by the Wall Street Journal ahead of a planned November IPO. The filing shows just how dramatically the Claude maker's business has scaled — and how expensive that scale has become. The San Francisco-based firm generated $16.3 billion in revenue during the first half of 2026, marking a dramatic surge from $1.1 billion in the same period last year. That growth came with a staggering cost: Anthropic reported a net loss of $80.7 billion for the first half of 2026, though $72 billion of that total stemmed from non-cash accounting adjustments related to convertible notes and financing instruments.

The infrastructure appetite behind those numbers is just as striking. Anthropic spent nearly $9 billion on computing capacity and infrastructure in the first six months of 2026 alone, surpassing its total infrastructure expenditure of $7.3 billion for the entirety of 2025. Longer term, the company's compute commitments have reportedly grown past the $600 billion mark, with additional deals under negotiation that could push total obligations toward $750 billion. Notably, the filing also acknowledges a dependency risk: approximately 20% of Anthropic's $16.3 billion first-half revenue was derived from just two customers, with the firm anticipating that a limited client base will continue to generate a significant portion of total revenue for the foreseeable future. It's a reminder that even the fastest-growing AI company carries real concentration and capital-intensity risk as it heads toward what could be a roughly $2 trillion valuation.

Nvidia bets on open models for the agent era

While labs chase capital, Nvidia used this week to extend its reach beyond chips and into the software layer powering autonomous agents. The Nemotron 3.5 Super VL is a large, 120-billion-parameter model geared for agents used by the enterprise. While it uses fewer parameters than the Nemotron 3 Ultra model, Nvidia says it can deliver similar accuracy in instruction-following and tool-use tasks.

Nvidia paired that enterprise-grade release with smaller, locally runnable models. The company also launched two lightweight Bolt models, intended to run locally on the new AI PCs and workstations running the DGX Spark — aimed at consumers, programmers, and enterprise knowledge workers, being released right now into the market through partners like Microsoft and Dell. Nvidia's pitch, in the words of its VP of Open Models, is that "any organization that wants to run agents on Nvidia infrastructure now has an open, optimized model for their tier without having to choose between capability and deployability." Early adopters reportedly include EY, Perplexity, CrowdStrike, Ramp, and Salesforce — a sign enterprises are eager to self-host agentic AI rather than rely solely on cloud APIs.

DeepSeek keeps the open-weight pressure on

Meanwhile, DeepSeek continued its rapid cadence of releases. Over the weekend it quietly pushed an updated version of its flagship efficient model, saying the update delivers "stronger benchmark performance, faster task completion, and shorter outputs than the previous version." Analysts noted that the rapid pace of DeepSeek's new launches underscores the intensifying contest between China's open-weight models and the proprietary models of their US counterparts, which are top performers, but with far higher costs. DeepSeek has also open-sourced the serving infrastructure behind that model, lowering the barrier for others to run comparable systems at scale — further intensifying competition on cost and openness rather than just raw capability.

Taken together, these three threads — Anthropic's eye-popping but precarious financials, Nvidia's push into agentic software, and DeepSeek's relentless open releases — suggest an industry simultaneously consolidating around a few capital-heavy giants and fragmenting toward cheaper, more accessible alternatives.

Sources

AI IPOAnthropicArtificial IntelligenceDaily AI UpdateDeepSeekNVIDIA