CVAI London: European Democracies on Alert, Bubble Worries, Anthropic Bullishness, Agents & Compute Shortages
Notes from talks with Cohere's Aidan Gomez, Alphabet's R. Martin Chavez, Recursive's Josh Tobin, Index's Danny Rimer, Sequoia's Luciana Lixandru & More
We just wrapped our mid-year Cerebral Valley AI Summit in London Wednesday. On stage, we heard from top model providers, AI application leaders, investors, and more.
We also anonymously surveyed the artificial intelligence insiders in attendance and you’ll definitely be surprised by what they had to say.
Key Takeaways from CVAI London
EUROPEAN SOVEREIGNTY. Both Cohere CEO Aidan Gomez and Sequoia Capital partner Luciana Lixandru saw opportunity in European leaders’ realization that they can no longer rely on the good will and protection of the United States of America. Gomez, whose company is Canadian, is in the process of merging his foundation model business with a German model provider. He’s living in Britain these days. “We want to build this alliance to help create capability beyond just one democracy,” Gomez said. Meanwhile, Sequoia just co-led a €500 million round in Stark Defense, a startup making kamikaze drones and other weapons. “Unfortunately, there’s nothing like conflict at the doorstep to really wake you up as a government,” Lixandru said. “Also I think governments realize that they’ve been under-investing in defense since before the war.”
MAJORITY WARN OF AI BUBBLE. They just don’t think it will pop this year. Fifty-seven CVAI London attendees filled out an anonymous survey that posed a series of questions about the AI industry today. (We’ll include the full survey results at the bottom of this post for paying Newcomer subscribers.) The majority of our AI insiders thought we were indeed in a bubble, but 51% said “yes bubble, not bursting this year,” compared with 5% who said “yes bubble, yes about to burst.” The remainder said, “no bubble, lots of room to run.”
ANTHROPIC BULLS. Multiple survey responses pointed to Anthropic optimism and OpenAI pessimism among the crowd. When asked what private unicorn they would like to own at today’s valuation, 54% said Anthropic. ElevenLabs was the second most popular answer with 13%. Surprisingly, Safe Superintelligence came in third with 8%. Meanwhile, when asked what company they would most like to short at its present valuation, a plurality of 33% answered OpenAI, with Perplexity next at 25%. Harvey came in third on the short list with 8%.
AGENT UTILITY & FORM FACTORS REMAIN TBD. An off-stage conversation with Stanislas Polu, co-founder of the agent orchestration startup Dust, illuminated some of the contradictions of agents at the moment. The company sees skills — specific, definable abilities — as perhaps the most useful framework for putting models to work. But many users prefer the concept of an agent with an identity and a particular set of skills. The industry is still sorting out whether humans will mostly interact with agents with unique identities and capabilities or with a more fluid, hive-mind-like AI with capabilities that can be turned on and off.
COMPUTE IS STILL HARD TO FIND. Recursive co-founder and CTO Josh Tobin was frank about one challenge his startup faces: finding enough compute. The lab is betting it can go all the way to recursive self-improvement with $650 million in funding co-led by Tom Hulme at GV and Greycroft, but it’s still having to run around the world pitching AI infrastructure providers on why they should let his startup pay for the privilege of using their hardware. “If it’s a bubble, there are no signs that I’ve seen that it’s like bursting right now,” Tobin said. “Basically all GPUs are essentially sold out, pretty much everywhere. People are kind of making reservations 6 to 18 months in advance and even for those it’s a competitive process.” Tobin argued that many AI foundation model companies believe that the compute shortage is only going to get worse. “One of things that’s fueling that dynamic is that there's a lot of people in the industry right now that believe that over the next couple of years the compute crunch is going to get even worse than it is today, and so it’s creating this race dynamic where these companies are buying up everything that they can get their hands on because they believe that the prices are going to go up in the future or availability is going to go down.”
FOUNDERS SHOULD GO DEEP ON EXPERTISE. The big labs may be incredibly well-capitalized and own the underlying model layer, but speakers on stage encouraged earlier-stage founders not to despair. While OpenAI and Anthropic pour gobs of money and research into AI coding tools, Tobin said that this focus is leaving audio, vision, video, and consumer applications genuinely underserved. ElevenLabs landed many plaudits throughout the day as one company which successfully carved out its market for voice models. R. Martin Chavez, the former Goldman Sachs chief information officer-turned vice chairman at Sixth Street, offered up the sage wisdom he got from Goldman’s head of sales: “Customers buy a product when they have unbearable pain and you have convinced them that only your software product can put an end to their pain. Everything else is just getting lucky.”
The Cerebral Valley AI Summit is co-hosted by Newcomer and Weekend co-founders Max Child and James Wilsterman.
Many thanks to our sponsors Nebius, Index Ventures, Higgsfield, and Weekend for making it possible.
Keep reading for the full rundown of the on-stage conversations.
A Transformer Paper Author Says We’ve Hit AGI
Aidan Gomez, CEO of Cohere and co-author on the famous “Attention Is All You Need” paper, opened the day by stating we’ve likely already reached AGI. “You can point it at any problem and it can basically exceed human capabilities,” he said.
General intelligence hasn’t come cheap, though. The price demands of token-maxxing are a short-term concern for coding tools, but Gomez says enterprise customers are still eager for more AI. “Everything else within the enterprise is from my perspective still essentially untouched.”
Gomez wasn’t very bullish on open-source models surpassing closed-source ones in the near future, because enterprises don’t want to be responsible for building and maintaining the infrastructure needed to run them.
He cautioned against the risk of the Western world relying too heavily on American foundation models. “A very small pool of these large, large tech players are becoming a single point of failure for the entire democratic block — and that is not a resilient system.”
Opportunity Shifts to the Infrastructure Layer and Hard Tech
Luciana Lixandru of Sequoia didn’t sugar coat the risk for AI application founders when OpenAI and Anthropic decide to come for their sectors. Sequoia has backed both, and she shouted out her partner Shaun Maguire’s investment in SpaceX as a strong bet through its merger with xAI.
That’s not to say there isn’t still enthusiasm for AI-native tools for specific verticals. She mentioned her firm’s investments Harvey for legal services, Sierra for customer support, OpenEvidence for healthcare.
But Lixandru said that she’s seeing fewer application companies get started than before, especially compared to those in infrastructure, inference, and robotics. She was careful not to be totally discouraging to anyone who wanted to go into vertical AI, though, as long as founders had “extraordinary domain expertise.” In Europe, she’s cautiously optimistic that the technical talent pool is starting to mature as more entrepreneurs have exited and are coming back as repeat founders.
Upgrading Legacy Businesses with AI Is a Massive Market
It’s easy to see that AI will be transformative for many large enterprises, but actually getting legacy businesses up to speed on new technology is still a big lift. Arnaud Fournier, CTO of Forward Deployed Engineering at OpenAI, is in the middle of such messy transition work, directly embedding company engineers in client businesses to demonstrate where the AI tools can be particularly useful.
The strategy, first pioneered by Palantir, comes with the advantage of troves of internal data that can be used to tailor OpenAI’s models to the specific needs of any company. “By having the data, owning the distribution, they have a head start implementing the technology,” he said.
One of his more prominent initiatives has been through the Thrive Holdings effort to integrate AI into accounting firm rollups. AI’s breakout use-case — coding assistants — has had surprisingly little salience there. “Accountants don’t code, and they won’t code with codex.”
Instead, AI has been very helpful for projects like document review, or for client intake for wealth managers who manage hundreds of clients. “I think a lot of the enterprise transformation we’re going to see is through non-coding tasks,” he noted.
Trust Is the Only Moat That Matters
Despite focusing on different parts of the AI stack, Eleanor Lightbody of Luminance and Neil Zeghidour from Gradium both agreed building AI products in legal tech and voice tech require earning a ton of trust, whether in a sensitive contract negotiation or a simpler service like a voice model that can pronounce URLs, telephone numbers, or postal codes correctly.
The best moat you can ask for as an AI founder is your customer’s trust, said Lightbody. Luminance started by only handling NDAs with its autonomous negotiation tools, for example, but advanced to build agents that can negotiate all kinds of contract terms after proving that its work was sound.
Once you build that trust, you gain a foothold over time by building deep internal expertise, and then it becomes possible to train models off of a company’s internal data to better personalize to their needs. Humans aren’t out of the loop yet in that fine-tuning process, said Zeghidour. “It’s a bit like optimizing cuisine with AI — it’s hard when they don’t have taste buds,” he joked.
Content Creation Is the Next Target for Agents
For Higgsfield’s Alex Mashrabov and Gamma’s Grant Lee, enterprise customers have still been the biggest drivers of growth even as they build prosumer tools. Higgsfield’s mix of enterprise customers jumped from 30% to 80% in the last 3 months, Mashrabov said.
Higgsfield has grown by generating commercial assets and video ad clips for customers. Agents and MCP interfaces have also been generating videos on their own, said Mashrabov.
Lee has been taking on a formidable incumbent in PowerPoint by quickly generating custom slide decks. Gamma will retain memory of your company’s brand guide, for example, and can also work very fast.
Both are optimistic that a lot of content creation for go-to-market teams can be fully automated in the near future. “You might envision a world where you’re sitting in the customer meeting, use Granola for all your notes, and the moment the meeting is done, you can actually have a presentation to send back to your client to say, ‘hey, these are all the things we discussed, these are the deliverables we agreed on, and these are the next steps,’” said Lee.
Plenty of Room for Multiple Winners
Index Ventures partner Danny Rimer pushed back on the idea that backing multiple labs is a contradiction. “We’re in the early innings,” he said, and the differentiation between the major models runs deeper than Pepsi versus Coke — it comes down to fundamental approach, the talent those approaches attract, and the culture that keeps that talent.
On Anthropic specifically, he was effusive: the clarity the team has about what it stands for, and its willingness to broadcast that and get into fights over it (including with the President of the United States) resonates both internally and with users.
On how founders should be building organizations right now, he credited Notion’s Ivan Zhao for a framing he found compelling: evaluate hires on capability, taste, and agency. Seniority, experience, and credentials matter less than ever.
The European startup ecosystem still isn’t pulling in as much funding as its stateside counterparts, certainly, and recent data suggests even Europe-based investors have been investing more in US founders than those closer to home. Rimer isn’t discouraged, though, and sees a lot of talent on the continent. He joked that those investors should leave for the states and leave all of the best deals for him.
AI Customer Service Goes Outbound
Customer support startup Decagon has made strong headway in the European market, and last year opened an office in London to be closer to its local clients. Decagon CEO Jesse Zhang said that counter to the common narratives, many of his European customers have embraced Decagon and rapidly adopted its products. “I think it’s because the board and the C-suite have top-down pressures to adopt AI. Our use case comes up quite often, so that has propelled a lot of the market to move quickly.”
Decagon has seen more success expanding into outbound agent calls, or “revenue-generating conversations.” Its customer Hertz, for example, saw a 25% reduction in unreturned cars and a huge uplift in rental extensions after the agents pre-emptively gave customers the option.
Zhang wasn’t too worried about the big labs eating into customer support, even as they develop cutting-edge voice models. “The model providers are building very wide and thin use cases. In the enterprise it’s a lot harder and there’s a lot of this last mile, extra tooling that’s very customer service-specific that you need to build.”
The New Frontier: Models Training Themselves
Josh Tobin and his Recursive Superintelligence co-founders left lucrative positions at OpenAI, Google, and Meta on the belief that recent AI breakthroughs have put us at an inflection point comparable to the shift from traditional machine learning to deep learning in 2015. Now what’s being transformed is AI research itself, where models are starting to train themselves.
For Tobin, recursive self-improvement hits when the primary innovations in a new model version come from the previous version of the model itself. That could mean AI researchers are eventually automated out of their own jobs — just as many fear software engineers could be with coding assistants.
On whether models can be genuinely creative, Tobin pointed to math proofs and to papers generated by early AI science systems that cleared peer review. He was also candid about what they see in testing: models escaping sandbox environments is “a real thing” that “happens a lot.” He was careful to separate that from existential risk, but didn’t dismiss it.
The White Collar Apocalypse Isn’t Coming Just Yet — But Workers Must Be Willing to Adapt
During the 2010s while still at Goldman Sachs, R. Martin Chavez saw the potential for AI agents from the moment he first pushed for “algorithmic trading.”
When he became CIO he told staff they had three choices — tell computers what to do, collaborate with the people who do, or stand in the way. If you picked number three, “that will be your last day at the company.”
But while he built a lot of the tech in-house decades ago, at his current firm, Sixth Street, he’s a buyer of software now rather than a builder. They use a company called Abacus, which he described as “a harness for models,” running the full investment process from first hearing about a company through generating the write-up that gets an investment committee to say yes.
He brushed off concerns that AI is going to replace white collar work in any major way. ‘There are way more people at Goldman after 15 years of bots — the business is way bigger, it makes way more money, does way more complicated trades, but if you were to list the activities of all the people 15 years ago and the people now, they’re completely different.”
See the Slides from Our Live Audience Survey
