SAN FRANCISCO — DeepSeek hit the wire this week, and the AI kings are sweating through their turtlenecks. The Chinese upstart claims it trained a high-performing model without the fanciest chips money can buy. Nobody in the Valley believed it. Now they're all saying the same word: amazing.
Call it a gut punch delivered on a budget. For two years the story out of Silicon Valley has been simple: bigger chips, bigger bills, bigger models. DeepSeek says it skipped the toll road and got there anyway, and engineers who've kicked the tires call it impressive even while working with less-advanced hardware than the U.S. labs hoard.
The reaction split the room. Some call it Chinese state-backed sleight of hand, numbers dressed up for the press. Others — the ones who actually ran the thing — say the results speak plain enough on their own, and that's the part keeping OpenAI and Anthropic executives up nights. If DeepSeek's math holds, the moat everybody built out of Nvidia chip orders just sprung a leak. Wall Street noticed too — the tech, media and telecom desks spent this week's market chatter on little else.
Meanwhile the American labs aren't standing still. Salesforce and Nvidia rolled out a new reasoning model called Koa, built atop Nvidia's open-weight Nemotron, trained to grind through sales calls, marketing copy and customer gripes without a human riding shotgun. Word from the trade is Koa does the unglamorous office work better and cheaper than the giant do-everything chatbots, which is exactly the kind of practical, boring competence that makes the fat contracts nervous.
Put the two stories on the same page and the picture gets clear fast. The expensive-chip, bigger-is-better playbook that built the American AI boom is getting squeezed from both ends — cheap and clever out of China, narrow and useful out of the enterprise shops. Trilogy's own shingles, from Aurea's CRM plumbing to the automation crews running through Skyvera's telecom stack, have built their whole business on doing more with less. This week the rest of the industry got a hard lesson in the same math.
Not every headline out of the AI beat was about chips and margins, though. LinkedIn co-founder Reid Hoffman put up $24.6 million to launch Manas AI, teaming with "The Emperor of All Maladies" author Siddhartha Mukherjee to point machine learning at cancer research. It's a reminder that while the big labs fight over silicon and market share, some of that same money is still chasing the older kind of return — the kind measured in years of life, not points of margin.
What happens next depends on who you ask on the trading floor. But the wire boys know a good scrap when they see one, and this week the scrap has a new fighter nobody trained for.
