![](\"https:\/\/fortune.com\/img-assets\/wp-content\/uploads\/2026\/04\/Julien-Bek-4-e1776793389629.jpeg?w=2048\")
<\/p>\n<\/p>\n
Hello and welcome to Eye on AI. In this edition\u2026Are services the new software?\u2026Anthropic\u2019s Mythos has financial regulators and bankers freaking out\u2026more executive turnover at OpenAI\u2026these measures may mean China will soon surpass the U.S. in developing the best AI models\u2026are AI inference costs getting too steep?<\/p>\n
Julien Bek never expected to go viral. Bek, who is an early stage investor in the London office of the venerable Silicon Valley venture capital firm Sequoia, says he merely wanted to highlight one of the firm\u2019s recent investing theses and use the piece to highlight some of the startups Sequoia had recently backed. So he penned a blog with the title \u201cServices: The New Software\u201d and posted it to his social feeds. Within days, it had surpassed 1 million views on X. It is now closing in on 3 million. It has done more than 450k impressions on LinkedIn.<\/p>\n
\u201cI certainly didn\u2019t expect to have this kind of reach,\u201d Bek told me on a call earlier this week.<\/p>\n
The provocative headline no doubt helped. But Bek\u2019s thesis also struck a nerve. In short, he thinks that the world\u2019s next $1 trillion company won\u2019t sell hardware or software as a product. Instead, it will sell an outcome, and use AI-powered software to help deliver it, alongside human expertise. Instead of selling customer service software, for instance, it will simply deliver customer service for a client, the way business outsource processing companies do today. But these new entrants will be AI-native from the start. Instead of selling legal tech, these firms will sell legal services, etc.<\/p>\n
Good examples of companies already pursuing this model that I\u2019ve written about before include both Robin AI and Legora in the legal space and Dwelly in the real estate market. There\u2019s also Dystyl AI in the consulting space, Rogo in financial services, and WithCoverage in the insurance brokerage market. Bek thinks there are many, many more to come. And he is sure the market potential is huge, noting that for every dollar enterprises spend on software, they spend six on services.<\/p>\n
Intelligence vs. judgment<\/p>\n
Bek has developed a taxonomy for thinking about these possibilities. First, he distinguishes between intelligence and judgment. Intelligence is basically anything with a pretty clear definition between the set of correct and incorrect answers\u2014think tasks in coding, mathematics, physics, and even some tasks in accounting, law, or medicine. AI models are getting pretty good at delivering intelligence. Judgment on the other hand is more about taste, professional intuition, and subtle but often critical qualitative distinctions that often require both talent and experience. Lots of companies are trying to figure out how to imbue AI models with judgment, but for the most part, they aren\u2019t there yet.<\/p>\n
He then performs a matrix analysis that plots how a given service ranks on an intelligence-judgment scale on one axis, and whether companies already tend to outsource a particular service, or perform it in-house, on the other axis. (This is a complex decision governed by economic ideas that Ronald Coase developed in the mid-20th Century and that I recently wrote about in the context of the so-called SaaSpocalypse for Fortune here.)<\/p>\n
First Bek looks at those tasks that companies already outsource to service providers, things like legal services, auditing, insurance brokerage, etc. Then he looks at the subset of those that are mostly about intelligence, with mainly just a dash of human professional judgment needed. This the sweet spot Bek thinks is ripe for AI-native service firms. \u201cIf [a customer] paid $100 for a service, but you offer them the same service for $80, but you can still do it at a high gross margin because you\u2019re using a lot of AI to deliver that service, then we think that\u2019s really interesting,\u201d he says. Among the functions he sees being in this category are things like insurance brokerages, insurance claims adjustment, IT managed services, tax advisory services, accounting and audit services, simple legal services, payroll services, and certain compliance services.<\/p>\n
Bek calls startups in this category\u2014heavy on intelligence, with a dash of judgment, in categories that customers already outsource\u2014\u201dautopilots.\u201d He says his use of that term in his viral essay has been the source of a lot of misunderstanding and misplaced criticism. He didn\u2019t use the term to mean that services could be performed entirely by AI agents to the exact same standard as human experts. What he meant was that the processes that deliver these services could be largely automated in the same way that autopilots function in aviation\u2014a human is still there monitoring the systems and handling the hardest tasks (like take off and landing) and ready to step in if something goes wrong, but a lot of the process is automated. He contrasts this to AI \u201ccopilots,\u201d where he says there is a lot more back and forth between the human expert and the AI system.<\/p>\n
I asked Bek about the theory that AI will enable some companies to in-source functions that they once outsourced. (That theory is part of what underlies the SaaSpocalypse\u2014the idea that companies will choose to make their own software using AI coding tools.) He allows that this may be true for some functions, but insists that there are many things that will never be in-sourced either because of regulatory requirements\u2014for example, financial auditing, in which companies must hire an independent firm\u2014or for what he calls \u201csofter\u201d reasons. The latter category includes things like management consulting, which exists in part to provide external validation of decisions management already wanted to take\u2014essentially helping to bolster their case to boards and investors, and, cynically, so that there is someone else to blame if it turns out to be a bad decision. The logic applies even in some IT functions. The old saying \u201cno one ever got fired for hiring IBM\u201d exists for a reason.<\/p>\n
Not just a lower bill, a different bill<\/p>\n
One of the biggest advantages the AI-native companies may have, Bek thinks, is around pricing. It\u2019s not just that the AI-powered service firms can potentially charge less, they can charge differently. Many services firms in many sectors have billed by time. Billing by outcome changes the game completely. \u201cWhen you\u2019re a smaller company, the best thing you can do to compete with the larger ones is actually disrupt them on pricing,\u201d he says. But he allows that it can take time to bring customers around to a different way of paying. For instance, people have been talking about getting rid of the billable hour in legal services for decades. The billable hour is, for most law firms that do corporate work, still here.<\/p>\n
Bek insists that there are signs the billable hour really is going away, in large part thanks to AI. (There was movement in this direction before AI, but AI certainly seems to have accelerated it.) But, at the same time, he acknowledges that impediments remain. Some large companies use RFPs from services like consulting that ask for \u201can hourly rate\u201d\u2014if you price differently, you might not get past that screening because you can\u2019t even complete the standardized form.<\/p>\n
What about margins? One reason investors have loved software businesses is because they have often been extremely high-margin. Once you create the product, you can replicate it and distribute it at almost zero marginal cost. Anything based on human labor doesn\u2019t scale the same way. Bek says the equation here is not as bad as some assume. In insurance broking, for instance, he says an AI-native startup like WithCover can sell 10x per human expert what a traditional insurance broker sells. \u201cSo I think the efficiency is proven, at least in some categories, not all,\u201d he says. \u201cBut I think this is very encouraging.\u201d<\/p>\n
Two costs that are a potential issue: the cost of AI inference and the go-to-market costs of selling a service. Inference costs for running AI agents can, in some cases, eat up a substantial sum of money. (More on that in the Brain Food section below.) Bek cites figures from Bret Taylor, the CEO of Sierra, which sells AI-based customer service solutions, that gross profit margins probably look like 70% instead of 90% for some pure SaaS companies. But 70% is still a healthy margin. But the go-to-market costs remain an unsolved challenge, Bek says. You can\u2019t scale enterprise service sales the same way you can software sales.<\/p>\n
Sequoia isn\u2019t the only investor with this idea. Private equity shops are betting that they can roll-up existing non-software businesses, infuse them with AI-driven efficiencies, and sell them off at much higher multiples. That\u2019s why OpenAI and Anthropic both have major sales channels being built around private equity firms. But Bek thinks AI-native startups will be able to grab substantial marketshare faster than legacy firms can metamorphize into AI-first orgs.<\/p>\n
He may be right. Change is hard. And having to reinvent both existing processes and existing business models is exponentially harder. The legacy companies have the relationships and the trust of existing customers. That\u2019s often a trump card, especially for the highest-value work. But at some point, delivering an outcome at a lower price might tempt many to at least try the AI-natives.<\/p>\n
With that, here\u2019s more AI news.<\/p>\n
Jeremy Kahn
jeremy.kahn@fortune.com
@jeremyakahn<\/p>\n
But before we get to the news: Do you want to learn more about how AI is likely to reshape your industry? Do you hear insights from some of tech\u2019s savviest executives and mingle with some of the best investors, thinkers, and builders in Silicon Valley and beyond? Do you like fly fishing or hiking? Well, then come join me and my fellow Fortune Tech co-chairs in Aspen, Colorado, for Fortune Brainstorm Tech, the year\u2019s best technology conference. And this year will be even more special because we are celebrating the 30th anniversary of the conference\u2019s founding. We will hear from CEOs such as Carol Tom\u00e9 from UPS, Snowflake CEO Sridhar Ramaswamy, Anduril CEO Brian Schmipf, Yahoo! CEO Jim Lanzone, and many more. There are AI aces like Boris Cherny, who heads Claude Code at Anthropic, and Sara Hooker, who is cofounder and CEO of Adaption Labs. And there are tech luminaries such as Steve Case and Meg Whitman. And you, of course! Apply to attend here.<\/p>\n
FORTUNE ON AI<\/p>\n
Anthropic\u2019s Mythos cybersecurity capabilities require urgent international cooperation, \u2018AI Godfather\u2019 Yoshua Bengio says\u2014by Beatrice Nolan<\/p>\n
Exclusive: Doctors and education experts who studied AI\u2019s impact on the young call for a 5-year moratorium in schools\u2014by Catherina Gioino<\/p>\n
Commentary: The hidden ROI of AI: What leaders should actually measure\u2014by Beena Ammananth and Jim Rowan<\/p>\n
AI IN THE NEWS<\/p>\n
Anthropic\u2019s Mythos model alarms international financial regulators. Global financial officials warned at the IMF and World Bank spring meetings that advanced AI models\u2014particularly Anthropic\u2019s Claude Mythos\u2014could pose systemic risks by rapidly uncovering and exploiting cyber vulnerabilities across banks and critical infrastructure. Policymakers including the Bank of England\u2019s Andrew Bailey and EU Central Bank chief Christine Lagarde said the technology could shift the balance between attackers and defenders, prompting calls for international coordination, though many regulators\u2014especially in Europe\u2014have yet to access or assess the model themselves. Read more from the Financial Times here.<\/p>\n
NSA using Anthropic\u2019s Mythos model despite supply chain risk designation. That\u2019s according to a scoop from Axios, which cited two anonymous sources. It said it was unclear how\u00a0 the National Security Agency was using the powerful new model, which Anthropic has only made available to a select few organizations to help them harden their cyber defenses. The model, according to Anthropic, possess unprecedented cyber capabilities, able to find zero day vulnerabilities in software code and string together multiple vulnerabilities into sophisticated autonomous attacks. (But the same methods can be used to point out flaws to defenders and help them patch them.) The use of Mythos though puts the NSA, which is part of the Department of War, and the U.S. government in an odd position since it recently designated Anthropic a \u2018supply chain risk\u2019 for insisting on a contract that would prohibit the U.S. military from using its AI models for lethal autonomous weapons or mass surveillance of U.S. citizens. Under the designation, the DoW and all of its contractors were supposed to stop using Anthropic\u2019s AI models.<\/p>\n
Anthropic CEO Dario Amodei holds White House talks viewed as effort at \u2018peace deal.\u2019 The CEO met with White House chief of staff Susie Wiles. It is unclear exactly what was discussed at the meeting and how much dealt with Mythos and how much with Anthropic\u2019s \u2018supply chain risk\u2019 designation, but the meeting was portrayed as an effort by the two sides to reach some kind of agreement to de-escalate their dispute and remove the supply chain risk label. You can read more here from The Washington Post. Meanwhile, Anthropic\u2019s legal challenge to that designation is heading towards a key May hearing in federal court in Washington, D.C., but at least some legal experts think the company\u2019s chances of having that particular three-judge panel overturn the designation aren\u2019t good. The same judges, who are mostly Trump appointees, refused to grant Anthropic a stay to prevent the ruling from taking effect. Anthropic won a separate challenge to the designation in federal court in California, but because the Pentagon used two different legal statutes to issue the designation, one of which can only be reviewed by the D.C. federal court, it remains in place.<\/p>\n
Amazon invests up to $25 billion into Anthropic. Amazon is investing an additional $5 billion in Anthropic, with the potential to commit up to $20 billion more tied to unspecified commercial milestones, expanding on its previous $8 billion stake, the companies said. As part of the deal, Anthropic plans to spend over $100 billion on Amazon Web Services over the next decade, using its infrastructure and Trainium chips while making its Claude platform more deeply integrated into AWS products. The agreement also secures Anthropic access to massive computing capacity\u2014including five gigawatts of AI compute and expanded inference resources across Asia and Europe\u2014as it continues partnerships with other providers like Google, Microsoft, and CoreWeave. Read more from CNBC here.<\/p>\n
Trio of senior OpenAI execs including Kevin Weil depart in further management shakeup. Senior OpenAI executive Kevin Weil, who was formerly one of Instagram\u2019s cofounders, is leaving the company, he announced last week. Weil had been vice president of product at OpenAI before moving over to a new AI for science division in October 2025. But now OpenAI is shuttering the Prism AI tool for scientific workflows that Weil\u2019s team developed, folding its capabilities into its Codex product. On the same day as Weil announced his departure, two other executives, Srinivas Narayanan, who had been CTO for OpenAI\u2019s B2B applications, and Bill Peebles, who had headed its now-discontinued Sora video-generation AI model team, also announced they were leaving. Their departures add to a wave of leadership turnover at OpenAI amid a wider restructuring that reflects the company\u2019s decision to focus on enterprise and coding products as it faces intensifying competition from Anthropic and prepares for a potential IPO.<\/p>\n
Startup from Google DeepMind, OpenAI alums valued at $4 billion just months after founding. The four-month old startup, called Recursive Superintelligence, raised at least $500 million at a $4 billion valuation, the Financial Times reported. Google\u2019s venture arm GV led the round with support from Nvidia. The company aims to develop a novel form of AI capable of continuously improving itself without human intervention. Its founders include Richard Socher, who founded genAI company You.com and previously led AI research at Salesforce, as well as Google DeepMind veteran Tim Rockt\u00e4schel, and former OpenAI researchers Jeff Clune, Josh Tobin and Tim Shi.<\/p>\n
EYE ON AI RESEARCH<\/p>\n
The U.S. is still leads in AI, but the trends are not favorable. That\u2019s one of the big takeaways from Stanford University\u2019s Human-Centered AI Institute\u2019s Annual AI Index, which dropped last week. (The Index is always a fantastic snapshot of where we are in AI development and contains so much information it would take many newsletter to summarize it all. Here I am just going to focus on this geopolitical point.) <\/p>\n
For years the U.S. dominated AI across nearly every meaningful dimension, but the new report documents how China has nearly closed that gap: U.S. and Chinese models traded places at the top of performance rankings multiple times during 2025, and as of March 2026 the leading U.S. model’s edge over its Chinese rival had shrunk to just 2.7%, according to Stanford. China now leads the U.S. in AI publication volume, citations, patent output, and industrial robot installations, while the U.S. still produces more top-tier models and higher-impact patents\u2014a distinction that may not hold for long given the talent trends. The number of AI scholars moving to the United States has dropped 89% since 2017, with that decline accelerating\u2014down 80% in the last year alone, as the Trump Administration has cracked down on both student and work visas.<\/p>\n
On investment the U.S. still vastly outspends China in disclosed private capital, but that comparison likely understates China’s total commitment, given its extensive use of government guidance funds estimated at $912 billion deployed across industries since 2000.The full picture, as ever, is more complicated than either the “America’s winning” or “China’s winning” narratives suggest, but certainly the U.S. may not be assured of having an edge in AI. You can read the rest in the full AI Index here.\u00a0<\/p>\n
AI CALENDAR<\/p>\n
April 23-27: International Conference on Learning Representations (ICLR), Rio de Janeiro, Brazil.<\/p>\n
April 22-24: Google Next, Las Vegas, Nevada.<\/p>\n
June 8-10: Fortune Brainstorm Tech, Aspen, Colo. Apply to attend here.<\/p>\n
June 17-20: VivaTech, Paris.<\/p>\n
July 6-11:\u00a0International Conference on Machine Learning (ICML), Seoul, South Korea.<\/p>\n
July 7-10:\u00a0AI for Good Summit, Geneva, Switzerland.<\/p>\n
BRAIN FOOD<\/p>\n
Are AI inference costs getting so steep that humans workers are a better deal? Only a few weeks ago, everyone was talking about \u201ctokenmaxxing\u201d\u2014developers competing with one another to use up all the available tokens in a particular license tier for top coding models such as Anthropic\u2019s Claude Code and OpenAI\u2019s Codex. (See this New York Times story on the phenomenon.)<\/p>\n
But, as ever in AI, a few weeks is a lifetime. Anthropic, experiencing a compute crunch, has capped the number of tokens users can consume on some pricing tiers during peak hours of the day. Meanwhile, its latest model, Claude Opus 4.7, also consumes more tokens and is more expensive to use, per query than its predecessors. OpenAI also recently changed its Codex pricing to charge users per token consumed as opposed to per message. The result is that some companies are finding their inference costs are soaring. So much so that one person posted to a Claude Code Reddit thread popular with developers that they had \u201cfired\u201d five AI coding agents and hired two mid-level human developers instead. (The downside, the person wrote, was that the company\u2019s coffee costs had now soared.)<\/p>\n
The post seems to have been intended as a joke. But it does reflect the feeling of a lot of developers after the recent price rises. Inference costs are also now a major pressure point\u2014one that may significantly slow AI diffusion across large enterprises.\u00a0<\/p>\n
#Sequoia #partner #thinks #AIenabled #services #software #Heres<\/p>\n","protected":false},"excerpt":{"rendered":"
Hello and welcome to Eye on AI. In this edition\u2026Are services the new software?\u2026Anthropic\u2019s Mythos…<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[245],"tags":[2473,2944,2211,8029,1862,410,9499,8780,2400,1123,9500,6924,2841,2401,3446,443],"_links":{"self":[{"href":"https:\/\/stock999.top\/index.php?rest_route=\/wp\/v2\/posts\/4491"}],"collection":[{"href":"https:\/\/stock999.top\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/stock999.top\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/stock999.top\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/stock999.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4491"}],"version-history":[{"count":0,"href":"https:\/\/stock999.top\/index.php?rest_route=\/wp\/v2\/posts\/4491\/revisions"}],"wp:attachment":[{"href":"https:\/\/stock999.top\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4491"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/stock999.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4491"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/stock999.top\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4491"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}