The dawn of the efficient AI model

Welcome to Runtime! Today: How DeepSeek's new AI model upended industry assumptions about the price of building leading-edge AI models, the U.K. will consider remedies to address cloud competition involving AWS and Microsoft, and the latest funding rounds in enterprise tech.

(Was this email forwarded to you? Sign up here to get Runtime each week.)

The beauty of constraints

When software engineers ran into roadblocks over the last few decades, they often resorted to a simple and time-honored solution: throwing more computing resources at the problem. That's how the first two years of the generative AI frenzy unfolded, with frequent chip shortages and Jensen Huang's evolution into the Elvis of AI, but DeepSeek just showed that there's another way to make progress.

DeepSeek's release of its R1 "reasoning" model last week forced everyone spending billions on AI infrastructure — under the premise that deploying massive amounts of computing power was the only way to improve AI model performance — to take a deep breath. The R1 model is on par with OpenAI and Anthropic's leading models despite using underpowered, much cheaper GPUs thanks to some novel engineering techniques, showing that "AI’s economics are being rewritten faster than anyone predicted," as Tomasz Tunguz of Theory Ventures put it.

• Wall Street took the news hard: Nvidia's stock fell 17% Monday and wiped out $600 billion in market cap, which CNBC said was "the biggest drop for any company on a single day in U.S. history."
• DeepSeek is a Chinese startup that was forced to train its models using Nvidia's H800 GPUs, a throttled version of the H100 chip that Nvidia designed to meet the Biden administration's chip embargo restrictions until it was also banned in late 2023.
• Those constraints forced DeepSeek's engineers to develop some clever workarounds, which Stratechery and The Next Platform unpack for those interested in the gory details.
• The end result was a world-class AI model produced using cheap hardware, although the total cost to train R1 was likely far more than the $6 million required to do the final training run, which was the number cited in initial reports that really set off the stock-market panic.

The R1 model was also released under the MIT license, which will allow other AI companies to replicate its approach and determine just how much the economics of AI actually did change. Hugging Face announced plans to do just that on Tuesday on a cluster of H100 chips, noting that it will face some challenges because DeepSeek didn't release any of the training data or assumptions behind its model.

• That process will help address the elephant in the room: no matter how cheap it is, U.S. enterprises are going to be very skeptical about putting an AI model designed in China at the heart of their tech stack.
• However, they're still interested in kicking the tires: in response to customer demand Snowflake added DeepSeek to its roster of models Monday, and AWS made it available through its Bedrock model service Tuesday.
• A version of R1 hosted by a U.S. vendor would be far more palatable: "All DeepSeek usage in Perplexity is through models hosted in data centers in the USA and Europe. DeepSeek is *open-source*. None of your data goes to China," Perplexity co-founder and CEO Aravind Srinivas said on LinkedIn Monday.
• And Redmonk's Stephen O'Grady noted that R1's cost structure could allow enterprises to deploy leading-edge AI models on their own self-managed infrastructure, which to date has been a non-starter for all but the richest corporations.

Nvidia's stock rebounded by nearly 9% Tuesday as the market started to realize that the techniques used to develop R1 could actually boost demand for AI in the long run by making it more accessible and affordable. Until just a few weeks ago, if you wanted to build cutting-edge AI models you needed huge amounts of capital and a close relationship with Nvidia, and when a wider group of people can start to experiment with new technologies interesting things tend to happen.

• a16z's Steven Sinofsky compared the breakthrough to the enormous changes that scale-out computing brought to infrastructure strategies in the mid-2000s, noting that "if history offers any advice to technologists, it is that core technologies become free/commodities and because of internet distribution and de facto market standardization at many layers that happens sooner with every turn of the crank."
• There are still a lot of caveats here, of course, chief among them proving R1's cost structure can be replicated at scale.
• But two years after Satya Nadella bragged about forcing the industry to "dance" to Microsoft's and OpenAI's strategy, DeepSeek just changed the tune.
• And Runtime can now safely recommend taking the under on the chances of Project Stargate spending $500 billion on data centers by 2029.

Duopoly, innit

After more than a year of deliberation, UK competition authorities have concluded that AWS and Microsoft exert too much power over cloud spending. Their "significant unilateral market power … harms competition in cloud services in the UK because it is harder for alternative cloud suppliers to enter and grow in these markets and customers face a limited choice of suppliers," the Competition and Markets Authority said in a statement, as reported by The Register.

The CMA board will now consider whether or not to implement remedies to address the situation, which could include forcing the Tenacious Two to interconnect their data centers or open up certain APIs. "Interventions to address some of these issues would allow UK businesses [to] get better deals from cloud providers, enabling them to contribute to economic growth," the CMA said, and a final decision is expected later this year.

Enterprise funding

Helion raised $425 million in Series F funding to build the world's first fusion energy plant, and cloud infrastructure providers desperate for new energy sources are closely watching its progress.

StackBlitz raised $105.5 million in new funding as it continues work on Bolt, an AI coding tool that also lets you deploy your apps from a browser interface.

Alice & Bob landed €100 million in Series B funding for its quantum-computing company, which was named after a key component in the original public-key cryptography paper.

Atomicwork scored $25 million in Series A funding to build a ServiceNow competitor in the IT service-management software market using AI agents.

Lanai launched with $10 million in seed funding for its AI management software, which is being built by enterprise software veterans Lexi Reese, Dr. Stephen Herrod, and Rajesh Raman.

Grepr launched with $9 million in seed funding to build an observability platform designed for engineers using, what else, AI.

The Runtime roundup

SAP beat forecasts for revenue and profit and raised its guidance for the full year, citing good progress among customers shifting from on-premises software to its cloud services.

NinjaOne announced plans to acquire Dropsuite for $252 million, hoping to add Dropsuite's data-protection software to its endpoint management and security platform.

Thanks for reading — see you Thursday!