DeepSeek Crosses $1B ARR, Prepping Shanghai IPO

DeepSeek Crosses $1B ARR, Prepping Shanghai IPO

The Information reported (via Reuters) that DeepSeek's annualized revenue run rate hit $1 billion as of September 24 — more than double where it stood just months ago [1]. CEO Liang Wenfeng shared the number directly with investors. The company is also finalizing a roughly $7.5B raise targeting a ¥500B valuation, with a Shanghai Stock Exchange listing in the pipeline [2].

The growth story is almost entirely API access, and the margins are the real headline: ~83% gross margin, with a meaningful chunk of the ARR jump coming from DeepSeek raising prices 2.3–4.5x [3]. That's not a company burning cash for share — that's pricing power in a market everyone assumed was a race to zero.

It also puts DeepSeek in the same conversation as MiniMax and Moonshot, both closing in on similar ARR milestones — a reminder that the "China can't monetize AI" narrative from early 2025 is aging badly. The center of gravity for frontier-adjacent model revenue is genuinely global now, not just a US story.

Snorkel AI Triples to $3.5B as Everyone Realizes Data Is the Bottleneck

Snorkel AI closed a $350M Series E at $3.5B — nearly 3x its prior valuation — led by Insight Partners and S32, with ARR reaching $350-375M, up 18x in a year [1][2]. The business: synthetic and complex training data, plus RL environments, sold directly to frontier labs [3].

This is the clearest data point yet that the bottleneck in frontier AI has shifted from compute to data quality — specifically the kind of complex, structured, RL-ready data that generic scraping can't produce. An 18x revenue jump in twelve months isn't hype-cycle froth; it's labs writing checks because they literally cannot train the next generation of models without better environments and labeled data at scale.

Founded out of Stanford's AI Lab in 2019, Snorkel expects profitability this year — a rare claim in this space and one worth taking seriously given the revenue trajectory. If data-as-a-service is where the margin lives now, expect a wave of specialized competitors chasing the same frontier-lab customer base.

Goldman Sachs: Data Center Power Demand Could Surge Up to 220% by 2030

Goldman Sachs' latest research pegs global data center power demand growing 165-220% by 2030 versus 2023 levels, driven almost entirely by AI, with global capacity forecasts around 217GW [1][2]. The US is expected to capture the largest share of that build-out [3].

The other half of the story: AI cloud providers are raising GPU instance prices 17-21% as demand outstrips supply, and hyperscalers are ramping capex accordingly. Pricing power on compute plus pricing power on frontier model APIs (see: DeepSeek above) suggests the "AI is getting cheaper" narrative is only true at the model-weights level — everything wrapped around inference is getting more expensive, not less.

For anyone modeling AI unit economics into 2027-2030 planning, this is the constraint to watch. Power and GPU access, not model capability, may end up being the actual ceiling on how fast agentic and simulation-heavy workloads (like Agora-2, above) can scale.