Deepseek just topped Ramp’s list of fastest-growing software vendors for June 2026, and that says more about where AI spending is headed than about any single Chinese lab. According to The Decoder, the category tracks breakout growth relative to a vendor’s size, and Deepseek led it outright. Ramp’s data pulls from real transactions across more than 50,000 companies, so this isn’t survey noise. US companies are paying Deepseek directly and routing data through its platform.
That last detail matters. Ramp chief economist Ara Kharazian points out this isn’t about teams quietly self-hosting an open-source model behind their own firewall. Companies are sending data straight to a Chinese provider. Kharazian flags the obvious security and competitive risks, and he doubts the trend holds. Worth noting: he’s the same economist whose index caught Deepseek’s first hype spike in January 2025, when adoption hit 0.3 percent before sliding back to 0.1 percent. He’s seen this movie before.
What’s Actually Driving It
Deepseek shipped V4 at the end of April. It doesn’t beat the best Western models on raw performance, but it costs a fraction of the price. Here’s the key line from the reporting: the performance gap is far smaller than the price gap. When a model gets you 90 percent of the way for 20 percent of the cost, finance teams start asking hard questions about whether the premium is worth it.
Kharazian reads cost awareness as the real driver behind Deepseek’s strong June. And it’s not an isolated signal. Inference platforms like Fireworks AI, fal AI, and DeepInfra are all growing as companies run open-source models instead of paying OpenAI or Anthropic per token. A December 2025 report showed Chinese models like Deepseek and Alibaba’s Qwen passed US rivals in Hugging Face downloads for the first time, taking over 44 percent of downloads for popular new models.
Why This Matters Now
Three forces are colliding at once:
- ROI pressure. Companies are struggling to measure AI’s return, and the debate over whether the spend pays off is getting louder.
- Rising prices. Model costs have climbed across every provider. The era of heavily subsidized flat rates is ending.
- A token economy taking shape. When buyers start picking models on price-to-performance instead of brand, the market starts behaving like a commodity exchange.
Put those together and you get exactly what Ramp is seeing. Buyers are treating intelligence as a metered utility, shopping for the cheapest acceptable token. The Decoder frames these as early signs, and that’s the right read. This is the start of a behavior shift, not the finish.
One myth the data kills: the so-called SaaSpocalypse, the idea that AI is wiping out established software. Ramp says design tools like Figma and Paper are still in demand even as Anthropic’s Claude spins off its own design product. Cheaper models aren’t torching the software stack. They’re reshaping how companies pay for the intelligence layer underneath it.
What Practitioners Should Do
The Western labs still lead by a wide margin, so don’t overread one breakout month. But the direction is clear, and there are concrete moves worth making now:
- Benchmark on price-to-performance, not brand. Run your actual workloads against cheaper models. The gap may be smaller than you assume.
- Separate the data question from the cost question. Sending sensitive data to a Chinese provider carries real risk. Self-hosting open weights or using a US inference platform gets you the cost savings without that exposure.
- Build for portability. An abstraction layer that lets you swap models is now a hedge, not a nice-to-have. Prices and rankings are moving fast.
- Watch the inference platforms. Fireworks, fal, and DeepInfra are where the token economy is forming. Their growth is a leading indicator.
Kharazian may be right that Deepseek’s specific lead fades. But the buying logic behind it, cost-first model selection, looks durable. Over the next year or two, expect more teams to route work to whatever clears the price-to-performance bar that day. The brand on the model matters less every quarter. You can find the full breakdown at The Decoder.