DeepSeek Heralds a New Phase in the AI Race: Low-Cost Competition
Plus, VCs are less in sync with Meta's VR ambitions
The Main Item
Tech Giants Hold Firm on AI Spending Plans, but DeepSeek Shows Upstarts Can Still Change the Game
OpenAI and other top AI companies succeeded last week in pushing back on the idea that China’s DeepSeek, with its low-cost solution, represented an existential threat to their businesses.
A blitz of product announcements from OpenAI and Google over the past two weeks showed they still have plenty up their sleeves, and Google, Microsoft, and Meta all affirmed in earnings reports that their massive capital spending plans remain intact. Amazon said it’s going even bigger.
But scratch a little, and there is deep uncertainty — along with no small amount of excitement — about what it all means for the industry and the world. Another low-cost LLM announcement, along with other developments this week, suggest that a reset of business models, along with even higher expectations of what AI can achieve, may be at hand.
One indication of where things are going came from researchers at Stanford and the University of Washington, who said they’ve developed a high-performance LLM for just $50 in cloud computing credits, according to TechCrunch. That makes DeepSeek look positively pricey, and suggests the Chinese company's innovations on lightweight models are just the beginning.
The effort relied on what’s called “distillation,” or using the output of one model to train another. OpenAI has said it’s investigating whether DeepSeek used its service in this way, which would violate its rules. Google’s Gemini similarly bars such use of its results.
The interwebs of course had a field day pointing out the irony of the big LLM companies complaining about others using their output, since their models are also built on the output of others (i.e. anyone doing anything on the internet.) While many media companies have cut deals allowing the LLM companies to crawl their content, The New York Times has not, and is going to the mat — even at a cost of almost $11 million and counting, as it revealed this week.
Even if the big proprietary models are ultimately able to bar distillation, though, the proliferation of high-performance open-source models could make any such restrictions moot.
Already, cheaper AI is putting price pressure on the AI ad-ons offered by enterprise software companies, notably Microsoft and Salesforce, according to The Information.
Power company executives, after enjoying massive run-ups in their stocks based on anticipated demand from AI data centers, are scratching their heads about what it all means for their very long-term investment decisions.
Market leader OpenAI, for its part, needs to show that it’s still ahead — and the new Deep Research product goes a fair way in accomplishing that, according to some early reviews. The company now has both an agent that can transact and one that can perform research that’s at least close to professional grade — accomplishments that might have stirred more buzz in a less dramatic week.
Efforts to ban DeepSeek as a security threat are now underway, though if the TikTok situation is any indicator, such concerns will ultimately take a back seat to business deals. With open-source LLMs proliferating, clever new approaches are sure to proliferate too, and will hardly be limited to China.
Lina Khan, the former FTC chair that big tech loves to hate, argued in an op-ed in The New York Times this week that DeepSeek’s achievement shows what’s possible in a more competitive environment. Left to their own devices, she says, big tech companies won’t push the frontier of innovation, lest it undermine their existing businesses.
She points out that Google researchers developed the “transformers” architecture at the heart of LLMs back in 2017, but did little with it for years.
Antitrust enforcement, along with China policy, is along the fault lines within Trumpworld, and especially among the president’s supporters in Silicon Valley.
Investors don’t like restrictions on big tech companies buying startups, but they also need a competitive landscape that gives their companies opportunities other than selling to a giant at the first opportunity.
The big tech vs. little tech debate is only going to get more heated, thanks to DeepSeek.
Brought to you by Vanta
Compliance for Startups: Download the Checklist
As a startup founder, finding product-market fit is your top priority.
But landing bigger customers requires SOC 2 or ISO 27001 compliance — a time-consuming process that pulls you away from building and shipping. That’s where Vanta comes in.
Join over 9,000 companies, including hundreds of Y Combinator-backed startups like Supabase, Newfront, and Fern who streamline compliance with Vanta’s automation and trusted network of security experts. Start with our SOC 2 compliance checklist, which breaks down the process into clear steps — so you can spend less time on compliance and more time growing your business.
One Big Chart
Meta’s Metaverse Plans Aren’t Matched From VCs
Meta may be locking its sights on DeepSeek in the open-source model race, but it still hasn’t given up on its vision for alternate reality tech.
The tech giant’s investment in its Reality Labs division topped $19.9 billion, per its latest annual report, and its overall investments in AR and VR have surpassed $80 billion since 2014. Meta is targeting more that $20 billion in spending to the sector for 2025, per the FT. Its Ray-Ban smart glasses sold over 1 million units last year.
Data from Crunchbase suggests that global funding to metaverse startups, broadly defined, peaked in 2021 at $6.24 billion. It’s now fallen to $1.48 billion as of 2024, an even smaller total relative to 2023’s measly $2.08 billion.
That being said, augmented reality and gaming are getting a boost from generative AI, VCs tell me.
One buzzy startup, Decart, landed a $32 million Series A back in December led by Benchmark for its playable “open world” model that generates new gameplay areas for its users as they’re playing.
Saga, which makes scalable blockchain infrastructure for apps, including game developers and AI agents, recently landed fresh funding from M13, Newcomer has learned.
Katelyn Donnelly, the managing partner of the edtech-focused fund Avalanche VC, sees virtual worlds as a ripe area for education applications, as long as they’re accessible on devices beyond headsets: “The problem with immersiveness is the hardware is expensive. Schools will buy it if they can amortize it over many students and many years.”
High prices appear to have limited Apple’s Vision Pro: The Information reported last fall that Apple is scaling back production of its headsets.
Meta’s early success with glasses, though, makes it an area to watch, especially as generative AI promises to speed up identification of objects in the world around us. “Billions of people wear either sunglasses or actual glasses in the world, so there’s enough scale there,” said M13 general partner Latif Peracha.
Several stealth startups have been building prototype devices in the category, said Lightspeed’s Moritz Baier-Lentz.
“Imagine having glasses on your face that aren’t particularly bulky that actually augment your world, logically or intelligently in a way that they help you navigate the world, remember people, classify things, learn or get information on things in your physical environment.”