February was about who pays for AI infrastructure, and whether the bill is about to get smaller.
The Spending
Microsoft set the tone at the start of the year. It said it expects to spend about $80 billion on AI-capable data centers in the fiscal year ending this June, with more than half of that in the US.¹ The earnings calls since then have pointed the same way: the largest tech companies plan to spend more on AI capacity this year, not less.
Then DeepSeek's R1 raised an uncomfortable question. If a strong reasoning model can be trained for a fraction of what US labs spend, does all that capacity pay off? My bet is that it does. Cheaper models mean more use cases clear the cost bar, and inference demand goes up.
Purpose-Built Hardware
NVIDIA's GPUs, including the H200, are still the default for large training and inference jobs.² A few companies are betting against the general-purpose GPU. Cerebras builds a single chip out of an entire silicon wafer. Its current WSE-3 has 4 trillion transistors.³ Groq and SambaNova focus on fast inference.
None of these will displace NVIDIA soon. They do give buyers real choices on latency and price, especially for serving models, where speed is what users feel.
Legal AI Raises Big
On February 12, Harvey raised $300 million at a $3 billion valuation in a round led by Sequoia.⁴ Harvey is a good example of a vertical AI company: one domain, deep integration into how lawyers already work, and customers who pay for accuracy. Anyone can call the same models Harvey uses. Owning the workflow is what earns the valuation.
The EU AI Act Starts
The first rules of the EU AI Act took effect on February 2. They ban a short list of practices, including manipulative AI and untargeted scraping of facial images. Within the week, the European Commission published guidelines on what counts as a prohibited practice and on what counts as an "AI system" in the first place.⁵
For product teams selling into Europe, the definition guidance matters most. It decides whether the Act applies to your product at all. The heavier rules for high-risk systems come later, so now is the time to figure out where you land.
What I'm Taking From February
- Infrastructure cost is a product decision. Latency and price per call decide which features are worth shipping.
- Plan for model costs to keep falling. Design features that get better as they do.
- Own a workflow. Vertical AI companies are raising at big valuations because the model is the easy part to copy.
- Read the EU definition guidelines now if you ship in Europe.
What infrastructure bets is your team making? I'd like to hear how other product people are thinking about cost and latency.