We recently met at Shopify’s Toronto office, where he described AI as “magical territory in which we teach sand to think.” In his words, “our computers now fulfill wishes.”

We discuss why pessimism sounds sophisticated and “optimism sounds naive,” how dystopian sci-fi shapes our expectations, and why agents could let people do things they couldn’t before. As he puts it, “dystopia is always just 30 years ago.”

There’s plenty of business in here, too: why Shopify welcomed Meta’s Muse when Amazon blocked it, how agents will make money, and the younger entrepreneurs he sees building businesses with AI and taking inspiration from MrBeast. We also discuss what founder control makes possible at companies like Shopify and Meta, where he sits on the board.

And, yes, I ask about the wine cellar in his house that he turned into a GPU server rack.

Sources is available wherever you get podcasts, including on YouTube, Spotify, and Apple Podcasts. Please subscribe and, if you feel so inclined, share the show with a friend.

This episode’s presenting sponsor is Hyperagent. Thanks to the show’s premiere sponsors: Atlassian, Granola, and Mercury.

Episode highlights

• Tobi’s advice to AI startups is to focus on the value their agents create before settling on a business model. I pressed him on the assumption that agents will collect a cut of every transaction they handle, and whose margin that money comes out of? He sees several possibilities, including charging users directly, metering usage like electricity, and taking a share of transactions. “If you build something that’s better than a cup of coffee, you are going to probably endure,” he says. That comparison felt familiar: when I priced this newsletter, I was thinking in cups of coffee, too. Advertising adds another complication: an agent paid to steer your purchases can become “a breach of trust,” but he still sees an opportunity for a new kind of advertising format for agents that maybe works more like a discovery marketplace.

• He says Shopify is seeing an explosion of new businesses, and the entrepreneurs are getting younger. “I think the new generation is extremely entrepreneurial,” he tells me. AI gives would-be founders a place to ask the questions they might once have emailed to a more experienced entrepreneur. But he also credits creators like MrBeast, whose young audience has watched him use money to build increasingly ambitious projects. The videos are entertainment, but Tobi sees an entrepreneurial education in how a creator reinvests in bigger ideas. For Shopify, he says he’d prefer to grow gross profit dollars by making entrepreneurship cheaper and more widespread than by taking a higher percentage from merchants. As he puts it, the goal is to give people “cruise control for their ambitions.”

• His viral GPU wine cellar is powering an AI assistant that helps him figure out who stresses him out. He didn’t have enough wine to fill the room, and buying more just to fill it didn’t feel like him. “I sometimes consult my inner 16-year-old,” he tells me. The servers serve a practical purpose, too: running his OpenClaw agent locally lets him use information he can’t send to an outside AI lab as a public-company CEO. He’s saved years of his writing and computer logs, and he’s also using calendar and heart-rate data to identify his most stressful meetings. He says he told the people involved, who didn’t find it as funny as he did. I suggested he check his heart rate after our interview, too.

ICYMI

Subscribe now

Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.

Even Google can’t get enough compute. I sit down with Google and Alphabet CEO Sundar Pichai shortly after the wider release of Gemini 4 Argon to discuss its early reception, where he still wants the model to improve, and what it takes to keep Google at the frontier of AI.

Pichai explains the recent leadership shakeup at Google DeepMind and why building frontier models has become an “all-consuming” job. We also get into Google’s answer to the personal agent race, how he predicts Gemini will do more work on your behalf, and whether agents will finally bring Search, Gmail, and Gemini together.

I ask him about Google turning free cash flow negative, the risks in the AI infrastructure buildout, and how he plans for demand years before knowing what the models will be capable of. He explains the signals giving him confidence to keep spending and how he uses AI agents to follow model development and compute usage inside Google.

We also discuss Project Suncatcher, Google’s new effort to put AI data centers in space, and why he thinks it belongs alongside bets like Waymo. He shares what’s next for Waymo and Wing, and when he expects quantum computing to become a real business.

Sources is available wherever you get podcasts, including on YouTube, Spotify, and Apple Podcasts.

This episode's presenting sponsor is Hyperagent. Thanks to the show’s premiere sponsors: Atlassian, Granola, and Mercury.

A MESSAGE FROM MY SPONSOR

Enterprise auth, now for agents

Enterprise buyers ask for SSO, SCIM, RBAC, and audit logs, and deals stall without them. Now they’re asking how agents get access too. WorkOS covers both: the enterprise features you already know, plus agent registration, scoped access, and OAuth 2.1 for your MCP server.

See how it works

Episode highlights

• Gemini 4’s early token usage is triple Gemini 3’s, but Google still needs to move faster. Pichai tells me Argon processed almost 25 trillion tokens in its first 72 hours, roughly three times what Gemini 3 handled in the same window last year. Even Google is running into capacity constraints, and he says more efficient versions are in development to expand access. When I ask how closely he follows the constant shifts in which lab is ahead, he says he pays attention to X but wants his teams thinking about being at the frontier a decade from now. “We didn’t get on this train recently,” he says. He also acknowledges that Google needed to pick up its release cadence this year and expects that faster pace to continue with Gemini 4.

• He says DeepMind’s leadership shakeup reflects how consuming the frontier model race has become. I ask him about Demis Hassabis stepping back from day-to-day leadership, Koray Kavukcuoglu taking over, and Jeff Dean leaving to start Discovery Loop. His explanation is that frontier model development has become an “all-consuming thing,” spanning data, compute, research, and managing both people and AI agents. He says Hassabis wants to spend more time on AI safety and applying the technology to science and healthcare, including through Isomorphic Labs, while remaining involved as DeepMind’s chair. Pichai describes Dean’s departure as a chance to pursue something new with people he wanted to work with. When I ask whether the changes mean DeepMind is being pulled closer to the rest of Google, he says that integration has already been happening for years.

• Google’s answer to Muse is a Gemini that keeps working after you stop chatting. When I ask about the personal agent race, Pichai says Google is bringing the capabilities behind its enterprise agents to consumers, including the ability to run tasks around the clock. His vision is “one box” that handles a conversation, a scheduled task, or work that needs to continue in the background. He says a growing share of Gemini requests already involve this kind of agent work and promises broader availability over the next few months. I also ask whether Search and Gemini will eventually merge. He says agents can provide a unified experience across Google’s products, carrying your context so you can start something in one place and pick it up in another.

• He defends Google’s spending even as free cash flow turns negative. He points to demand across consumers and businesses, improving model capabilities, and long-term contracts with committed revenue. “It’s not just hope,” he says. The planning problem is that Google is living with compute decisions made two years ago while being judged against what everyone knows today. I push him on the fragility of AI infrastructure financing and whether the broader market is getting ahead of itself.

• He sees data centers in space as a serious answer to Google’s long-term compute needs. I ask why Google is investing in Project Suncatcher when there are still deserts, oceans, and other places on Earth where it could build. Pichai points to the prospect of constant solar power in orbit and the constraints of finding room for decades of growing compute demand on the ground. He compares the project to Waymo’s early days, when imagining what it would become sounded like science fiction, and talks about a 10-to-20-year horizon. He says the first Suncatcher pilot is doing well. When we taped on Monday, he teased a “hello world” moment coming soon, when the world would get to “say hello to the TPUs.”

ICYMI

Subscribe now

Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.

Sources is available wherever you get podcasts, including video or audio on YouTube and Spotify, and audio on Apple Podcasts.

Thanks to the show’s premiere sponsors: Atlassian, Granola, and Mercury.

Episode highlights

• Tenev doesn’t think giving everyone AI will eliminate the need for a trading edge. If everyone uses the same AI tools in the same way, there’s no advantage. He sees differentiation coming from the data people give their agents, the tools they connect, and strategies that combine multiple asset classes. Robinhood’s bet is that agents can handle complexity that would overwhelm an individual trader. But access to a powerful model is only the starting point: “You’ll have to figure out how to be different.”

• He thinks part of the AI backlash comes from people having no ownership in the companies driving it. He contrasts AI with crypto, where individuals could invest early and become a constituency willing to defend the technology. With AI, many people see resource-hungry data centers making a small group of investors richer while they get none of the upside. That’s his case for Robinhood Ventures and opening private-company investing to everyday investors: people are more likely to support an industry’s growth when they have a stake in its success.

• Robinhood could give everyday traders access to one of Wall Street’s most physical advantages: being closer to the exchange. He floats letting users deploy trading strategies on infrastructure co-located with an exchange’s matching engine, reducing the delay between sending an order and having it processed. That proximity helps sophisticated trading firms run strategies others can’t. Robinhood doesn’t offer it today, and he gives no timeline, but he wants to explore making it accessible without requiring data center or networking expertise. “Can we democratize that?”

ICYMI

Subscribe now

Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.

AI could give people a reason to buy a new phone again. Qualcomm CEO Cristiano Amon tells me why he expects a smartphone supercycle as AI agents start using our apps and acting on our behalf. His company makes the chips behind Android phones, Meta’s smart glasses, and coming devices from AI companies like OpenAI, giving it a stake in several competing visions for the next big AI device.

We discuss why he thinks phones aren’t going away, how agents will use our personal context, and why he expects smart glasses to become the biggest category of AI wearables. We also cover differences in consumer AI adoption between China and the US, Qualcomm’s acquisition of Modular to offer developers an alternative to Nvidia’s software, and what 6G could change about the devices we use.

This conversation was recorded live at Qualcomm’s recent Snapdragon Summit in Hawaii.

Sources is available wherever you get podcasts, including video or audio on YouTube and Spotify, and audio on Apple Podcasts.

Thanks to the show’s premiere sponsors: Atlassian, Granola, and Mercury.

A MESSAGE FROM MY SPONSOR

Reliable Clean Energy

Constellation is America’s largest producer of reliable, clean energy, driving progress and innovation coast to coast. From local neighborhoods to iconic cities, we bring the energy that serves industries, strengthens communities, and moves America forward.

Learn More

Episode highlights

• A reason to upgrade your phone again. Amon thinks AI agents that use your apps, understand your personal context, and act on your behalf will drive the next smartphone supercycle. He compares this moment to the early iPhone, when the applications that would define the platform mostly hadn’t been invented yet. “The smartphone was defined by the third-party applications,” he says, and he expects agents to change the phone experience in a similar way. He says AI companies are already asking Qualcomm to support at least 100-billion-parameter models on phones by 2028, with AI running all the time. He bets these new capabilities will meet pent-up demand from buyers who have been putting off upgrades because of soaring memory prices.

• An “iPod moment” for AI devices. He describes companies experimenting with earbuds with cameras, jewelry, pendants, and badges, with Qualcomm supplying the technology behind a range of designs. He’s most bullish on smart glasses, which put an AI assistant close to your eyes, ears, and mouth. But when I ask whether another form factor could break through alongside phones and glasses, he sees room for something we haven’t settled on yet. “It feels like an iPod moment,” he tells me. In his view, the conditions for a hit are straightforward: people have to trust it, find it useful, and be able to use it without much friction.

• Qualcomm wants its mobile DNA to pay off in data centers. He thinks years of squeezing more computing out of limited battery power give Qualcomm an opening as AI infrastructure runs up against energy constraints. “Our DNA is building chips, assuming there’s a battery on the other side,” he says. The hardware bet is energy-efficient CPUs and AI accelerators; the acquisition of Chris Lattner’s Modular is meant to give developers an open software platform that works across competing chips. Amon says Qualcomm will “celebrate every single Modular deployment on Nvidia,” because loosening software lock-in gives it a chance to compete on the merits of its hardware. He sees 6G enabling constant high-definition video uploads from devices like glasses, turning us into “walking cameras on this world” that provide context for AI. He also describes using the network’s radio signals like radar to map the physical world for applications including robots and drone detection. That future will take time: he expects demonstrations at the 2028 LA Olympics, early deployments around 2029–2030, and broader scale after that.

ICYMI

Subscribe now

Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.

Two years and $4.6 billion in funding later, a startup called Reflection is soon releasing its first open-weight model in a bid to become the DeepSeek of the West.

Reflection’s new model, Beam, is arriving at an interesting time: Chinese labs have been setting the pace at the top end of open models for a while, and the Western world has lacked an answer. As more workloads shift to open models, who develops them matters more. A competitive Western alternative to Chinese open models would give businesses and governments another serious option for AI they can run and customize themselves.

Reflection’s founders have helped build some of the systems that got us here. CEO Misha Laskin led reward modeling for Gemini at Google DeepMind. CTO Ioannis Antonoglou helped create AlphaGo and AlphaZero, then led Gemini’s reinforcement learning from human feedback effort. Nvidia, Sequoia, and Lightspeed are among the investors backing Reflection, which was last valued at $25 billion.

Beam was trained from scratch and has 501 billion parameters. Reflection says early tests put it alongside leading open models on coding and agent tasks, and that it uses roughly a quarter of GLM-5.3’s inference compute in its reasoning-efficiency comparison. An Apache 2.0 license also means developers can modify it and build businesses on it without special commercial restrictions. “Closed models are the equivalent in real estate to renting an apartment,” Laskin says. Now, he argues that “an ownership market” for AI is starting to form.

The big question is how Reflection makes money while giving Beam away. We get into that, its plans to help companies and countries run their own AI, and why the founders think a handful of closed labs shouldn’t control where the technology goes. I also press them on safety and whether they'd consider a future model too powerful to release.

Sources is available wherever you get podcasts, including video or audio on YouTube and Spotify, and audio on Apple Podcasts.

Thanks to the show’s premiere sponsors: Atlassian, Granola, and Mercury.

A MESSAGE FROM MY SPONSOR

Powering Innovation

Constellation is America’s largest producer of reliable, clean energy, driving progress and innovation coast to coast. From local neighborhoods to iconic cities, we bring the energy that serves industries, strengthens communities, and moves America forward.

Learn More

Episode highlights

• Reflection trained Beam from scratch, without distilling another lab’s model, because it wanted control over the entire training process. Antonoglou says owning the architecture, data, and pretraining was essential to pushing reinforcement learning further. The initial training had to give Beam the reasoning capabilities that reinforcement learning could then build on. That control also helped the team keep the model stable as it poured more compute into training. There was also a geographic consideration: Reflection wanted its teams in the US and UK to develop a Western model end-to-end.

• Reflection’s business is selling the software and infrastructure that make its open models useful to companies and governments. Customers need help deploying agents, customizing models, and running them at scale, which is where Reflection expects to make money. Its $1 billion compute deal with Nebius and multibillion-dollar agreement with SpaceX give a sense of the scale of that bet. Over the long term, Laskin says it “probably makes sense” for Reflection to become its own hyperscaler, allowing Reflection to capture more margin. For now, the priority is getting its models out and building demand for running them.

• Reflection’s founders see concentrating AI in a handful of closed labs as a safety risk. Laskin argues that a small group of researchers cannot find every vulnerability, while open models let a much larger community inspect the technology and develop safeguards. Antonoglou says putting so much power inside one or two institutions creates “a single node of failure,” even when those institutions have good intentions. Their preferred outcome is a mix of open and closed systems, with different providers able to help defend against failures elsewhere. When I press them on whether a future Reflection model could be too powerful to release freely, they acknowledge that greater capabilities could justify restrictions. Laskin says those decisions should turn on what a model can do: “That is independent of whether they are open or not.”

ICYMI

Subscribe now

Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.

OpenAI’s next big bet is Dots, personal AI agents that can find ways to help you without waiting for a prompt.

Alexander Embiricos, OpenAI’s Dots product lead, tells me why the company is moving beyond chatbots and what that means for the future of ChatGPT. We discuss why Dots have their own computers, how they share context across the places you talk to…

Factory used to lose engineering recruits over a sentence on its website: “The future is IDE-free.”

In 2023, the company was betting that developers would eventually move beyond the coding tools their work revolved around. Engineers interviewing to join thought the idea was insane.

“We almost took it down,” CEO Matan Grinberg recently told me, referring to the IDE-free slogan.

Now, Factory is ascendant, having just raised an additional $200 million at a $5 billion valuation. While the first wave of AI coding tools helped engineers write the next line, Factory has been building for a world where they can hand off the entire task, from start to finish, to the company’s Droid agents. As Grinberg explains, Factory is well positioned for coding work to spread across different models. The bear case for Factory is that one model wins everything.

We also discuss why he thinks AGI is already here, how Factory hires its own engineers, his view of the open-versus-closed AI debate, competition with companies like Cognition, and when companies should outsource to AI rather than build in-house.

Sources is available wherever you get podcasts, including video or audio on YouTube and Spotify, and audio on Apple Podcasts.

Thanks to the show’s premiere sponsors: Atlassian, Granola, and Mercury.

A MESSAGE FROM MY SPONSOR

Books that keep pace with you

You shouldn’t have to wait for next month to know where your business stands today. Introducing Mercury Books, intelligent accounting built into Mercury for faster closes, real-time reporting, and always-on clarity.

Explore Mercury Books

Episode highlights

• Matan expects more people to do engineering, even if they don’t call themselves engineers. His analogy is the iPhone: taking thousands of photos no longer makes someone a photographer. As AI handles more of the coding, he sees the valuable skill as understanding a problem and defining what a solution needs.

• He’s building for a world where people want many different models. Some tasks need a cheaper, faster model; others can cost less if you’re willing to wait. Matan argues that software should handle those trade-offs between capability, speed, and token costs. He compares it to making toast: “You don’t give a shit where the electricity in your toaster came from.”

• He thinks we’re already living in a “post-AGI world.” Matan’s argument is that capabilities that would have stunned us four years ago now feel routine because we adjust so quickly. After a week of surprise at the capabilities of a new frontier model, he says, the reaction becomes: “All right, table stakes now. What’s next?”

Subscribe now

ICYMI

Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.

I recently sat down with Microsoft CEO Satya Nadella in Seattle for the unveiling of the new Copilot.

We discuss Autopilot, Microsoft’s new OpenClaw-based agent that can work on your behalf, and why he thinks newer AI models are finally capable of delivering on more of Copilot’s promise.

Nadella also has a blunt assessment of the AI industry: “We are way too self-obsessed.” I ask him why the industry has struggled to explain its benefits and what it will take to earn people’s trust. We get into his concerns about agents acting deceptively, when a safety problem should stop a release, and why he doesn’t want AI oversight to become a “cartel-like arrangement.”

We also discuss Microsoft’s relationship with OpenAI, the models Microsoft is building itself, and why Nadella wants investors to think about its apps, agents, and infrastructure as one connected business. He explains how he uses AI to track Microsoft’s capital spending, shares his vision for “unmetered intelligence” on Windows, and gives an update on Xbox’s path back to growth.

Sources is available wherever you get podcasts, including video or audio on YouTube and Spotify, and audio on Apple Podcasts.

Thanks to the show’s premiere sponsors: Atlassian, Granola, and Mercury.

A MESSAGE FROM MY SPONSOR

AI shouldn’t dictate where your models run.

Run the models you want, securely, wherever they make the most sense. Across clouds, data centers and the edge, VAST Data is extending its AI Operating System to give organizations control over model choice, placement, access, and cost.

See what’s possible.

Nadella highlights

• He wants to bring Autopilot’s AI chief of staff to your personal life, too. Built on a hardened version of OpenClaw, the agent gets its own computer, workspace, and memory to work continuously. When I ask about Microsoft’s consumer strategy, he points to its 100 million-plus consumer subscribers: “Autopilot should also go to the consumer side.” He sees consumer agents potentially cutting out today’s middlemen, making that market more zero-sum. In the enterprise, he thinks agents will make the market bigger than cloud “by orders of magnitude.”

• He thinks the AI industry is too self-obsessed to explain why anyone else should care. “I think we are way too self-obsessed as an industry about, look at us, how glorious we are,” he tells me. His prescription is to let the people using AI speak to its benefits and demonstrate that it creates economic opportunity for workers and the communities hosting data centers. More pronouncements from tech CEOs won’t do it: “Any amount that I say or any one of us say is not good enough anymore, I feel.”

• Microsoft wants its own frontier models and to serve all the others. Nadella says Microsoft’s models are already being used in Copilot’s Auto mode, and its frontier ambitions are tied to its own data and customer needs. He also makes the economic case for competition: “You can’t have one model,” he says, warning about where token pricing would end up. Microsoft intends to keep supporting other labs as it develops its own models: “Copilot will still have all the other models, because customers will expect that.”

Subscribe now

ICYMI

Sources is a podcast and newsletter by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.

For my first sit-down with Salesforce CEO Marc Benioff, I had to know one thing: is the framemogging intentional?

He tells me it’s not, but he also seems to enjoy the memes he creates every year from Dreamforce, the multi-day Salesforce bonanza that takes over San Francisco each September. At this year’s conference, he took over social media by towering beside Sam Altman, Dario Amodei, and Jensen Huang onstage.

Of course, Benioff has a lot more going on than generating memes. After a brutal, SaaSpocalypse-fueled drawdown, Salesforce’s stock price is up more than 20% in the last six months. Slack is “up into the right in a way that we haven’t really seen,” he tells me. We also discuss the resurgence of San Francisco, his role in the recent acquisition of Hugging Face, his ownership of TIME magazine, the new wealth being created by the upcoming IPOs of Anthropic and OpenAI, and what he thinks the AI industry can learn from the social media era.

Sources is available wherever you get podcasts, including video or audio on YouTube and Spotify, and audio on Apple Podcasts.

Thanks to the show’s premiere sponsors: Atlassian, Granola, and Mercury.

A MESSAGE FROM MY SPONSOR

AI shouldn’t dictate where your models run.

Run the models you want, securely, wherever they make the most sense. Across clouds, data centers and the edge, VAST Data is extending its AI Operating System to give organizations control over model choice, placement, access, and cost.

See what’s possible.

Benioff highlights

• He says Salesforce bought Slack because it saw the future of AI agents coming. “Our premise was that it was going to be the agent for business,” he told me. When I questioned whether that was really the thinking six years ago, he said he still has the paper written by Salesforce’s chief futurist, Peter Schwartz. Benioff acknowledged paying a steep price but argued that Slack’s role inside companies like OpenAI and Anthropic has vindicated the deal: “It was going to trade once, and we felt that we needed to step up.”

• Salesforce’s Anthropic bet started with being shut out of OpenAI. Benioff told me Microsoft’s relationship with OpenAI prevented Salesforce from investing years ago, sending him toward rival model companies. Salesforce initially invested roughly $50 million in Anthropic in 2023, when its Series C was priced at a reported $4.1 billion pre-money valuation, and has since invested more than $300 million across multiple rounds. He predicts Salesforce could make tens of billions in the Anthropic IPO. That financial relationship is now deeply embedded in the products: their recent Claudeforce partnership makes Claude the default model in Slack, including Slackbot, and brings Salesforce sales workflows directly into Claude.

• Benioff thinks the next big AI shift is an interface that changes around you. He pointed to my iPad during the conversation, noting that its icons and email list hadn’t changed. His prediction is that the interface will listen, adapt, and proactively surface useful information, even suggesting a follow-up question based on something he’d just said and my previous interviews. “That is not going to be true for very much longer,” he said of today’s static screens. He expects software to feel “live or living,” and thinks the coming interface change is the part of the AI story people aren’t yet appreciating.

ICYMI

Subscribe now

Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.

Lately, I’ve been spending a lot of time thinking about the AI inference market and how big it could get. Agents like Muse, Instinct, Town, and Grok Bot aggressively use browsers, run their own computers, and burn through far more tokens than a traditional chatbot. As more people put them to work (Muse is number two in the App Store), the demand for inference, or the computing needed to run these models, should grow enormously.

That’s why I was excited to talk with Tuhin Srivastava, the CEO and co-founder of Baseten, for episode three of the Sources podcast. His company builds the infrastructure that AI applications use to serve their customers. You may never have heard of Baseten, but Tuhin says that if you use popular AI apps, there’s a good chance some of your AI activity already runs through his platform.

Baseten’s growth gives a sense of what’s happening: The company raised funding at a $5 billion valuation in January and again at $13 billion in June. Tuhin tells me its token volume has grown 40 times year over year, with revenue up roughly tenfold over the last 12 months. He expects Baseten to become one of the largest users of compute on the planet within the next year or two.

Of course, growing demand doesn’t guarantee that an independent company like Baseten wins. It’s competing in a market the big clouds also want, where access to GPUs alone is a hard business to defend. I ask Tuhin what makes Baseten more than a reseller of someone else’s chips and why he thinks companies will increasingly want to own and improve their AI using their own data.

We also discuss Baseten’s recent acquisition to power agents working in browsers and virtual machines, why companies are embracing Chinese open models, and his ambitions for Baseten’s research lab. After my conversations with Sam Altman and Mark Zuckerberg, I wanted to hear from someone building the infrastructure behind this shift. Tuhin is betting that we’ve barely started to see how much of it we’ll need.

In addition to Substack, the Sources podcast is available wherever you get podcasts, including audio on Apple Podcasts and video on YouTube and Spotify.

Thanks to the show’s premiere sponsors: Atlassian, Granola, and Mercury.

A MESSAGE FROM MY SPONSOR

Here’s something for busy people like you

When you’re back-to-back all day, there’s no time to process. No time to write the follow-up. No time to turn “we should probably...” into, “ok done”.

Granola helps you become the person who actually does what they promised they’d do.
You can still make small notes in the meeting, or not, either way Granola quietly transcribes in the background, then turns those notes into clear summaries with actual next steps. After your call, share the notes with the team so everyone’s aligned. Or even chat with the actual notes to pull out exactly what you need to do next. No need to re-read the whole thing.

Download Granola and try it in your next meeting.

Episode highlights

• Tuhin says the AI industry has done a “pretty terrible job” bringing the public along. The backlash against data centers affects Baseten’s ability to get the compute it needs, and he acknowledges that communities have valid concerns. In today’s episode, I press him on how the industry can explain AI’s benefits while its own researchers warn about the risks.

• He thinks inference could be the only market left after AGI. His argument is that even if models become capable of doing everything, they’ll still need computing power and software to do it. “Without AGI, inference is going to be the biggest market ever,” he tells me. With AGI, “it’ll be the only market left.”

• Baseten’s defense against becoming a thin GPU reseller is to help companies own their AI. Tuhin describes a “continual learning loop” in which customers run models, evaluate the results, improve them with their own data, and repeat. His bet is that helping companies control that process and protect their margins will make Baseten harder to replace.

• AI agents are already deploying models on Baseten. Tuhin says agents are choosing which models to use and running them through the platform. Baseten also uses agents to investigate customer incidents, though it hasn’t given them free rein to make consequential changes in production. When I ask what he’ll never delegate to an agent, he says spending time with other people. “All we ever had,” he tells me, “is people.”

Subscribe now

Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.

Read more

When Mark Zuckerberg laid out his “personal superintelligence” vision over a year ago, he was thinking about what Meta released today: Muse, an AI agent with its own computer that works for you around the clock. His ambition, he tells me, is to “give every person in the world a very capable personal agent that can understand their goals and can just work on their behalf 24/7.”

I sat down with Zuckerberg at Meta headquarters last Friday, and I’ve been trying Muse for the last several days. I’m impressed. It’s essentially OpenClaw for normies, with a serious focus on security that I think will make this kind of product accessible to a lot more people. You can also connect Muse to your Instagram and Facebook accounts, including the tools for running a business and buying ads.

Zuckerberg’s ambition is to build something he can hand to anyone in his family, regardless of their technical ability. “I don’t think there are going to be billions of people who are going to buy a Mac Studio and configure it and run it at their home,” he tells me.

Available now everywhere you get podcasts.

Getting people comfortable giving an agent access to their lives is another challenge. Zuckerberg tells me he personally helped recruit the founder of Signal to work on a confidential virtual machine designed to keep even Meta from seeing what Muse is doing on a user's behalf. That’s an important promise from a company asking people to connect their email, messages, and other personal information to an AI that can act on their behalf. We get into how Meta is approaching that problem and why Zuckerberg sees privacy as a reason people will choose Muse.

He’s also making it very hard to compete on price. Muse launches with up to 100 million tokens per week for free, enough to give people plenty of room to put an agent to work. Zuckerberg thinks Muse will eventually pay for itself by helping people earn and save money, with Meta taking a small cut of transactions and offering a paid subscription for power users. We discuss how he uses Muse himself, from organizing baking projects with his daughter to critiquing his MMA training, and what happens when Muse agents start interacting with one another.

We also get into the backlash against data centers, why he thinks speculative developers are making it worse, Meta’s youth safety settlement, and the privacy concerns surrounding its smart glasses. He explains what went wrong with Llama 4, how he rebuilt Meta’s AI lab around him, and why he disagrees with rivals keeping their most powerful models to themselves.

A MESSAGE FROM MY SPONSOR

Reliable Clean Energy

Constellation is America’s largest producer of reliable, clean energy, driving progress and innovation coast to coast. From local neighborhoods to iconic cities, we bring the energy that serves industries, strengthens communities, and moves America forward.

Learn More

Episode highlights

• Zuckerberg is giving away a lot of compute to win the personal agent market. Muse launches with up to 100 million tokens a week for free, plus a virtual machine to work on your behalf. With startups like Instinct and Town chasing the same market, Meta is setting an aggressive expectation for how much a personal assistant should do before asking you to pay. Those startups don’t have Meta’s existing business to subsidize that usage. Zuckerberg’s plan is to eventually “take a very small cut of whatever the transaction is,” potentially paid by the businesses users transact with. “We think that this thing is actually going to make you money and save you money, and that is how it’s going to pay for itself.”

• Zuckerberg personally helped recruit Signal’s founder to make Muse’s virtual computer private. He says he and Nat Friedman recruited Moxie Marlinspike to work on the confidential virtual machine behind Muse. The ambition is to give people a computer in the cloud with the confidentiality they’d expect from a machine under their own desk: “Even Meta cannot see the content that is in there.” Zuckerberg says he isn’t aware of anyone offering anything close to that system. He also describes separate “sentinel” agents that monitor what Muse receives and sends, flag possible prompt injections, and require user approval for sensitive actions.

• He thinks speculative data center developers are helping fuel the backlash against AI infrastructure. Zuckerberg draws a distinction between companies like Meta, which he says plan to invest in communities for decades, and developers trying to secure a site and flip it. “They don’t really care as much about the local community, and they’re not invested for the long term,” he says. He argues that Meta’s projects face less opposition because the company has a reason to invest in local jobs and training.

• He responds to Meta’s landmark youth-safety settlement and to privacy concerns over the company’s smart glasses. Zuckerberg says he wants the youth safety settlement to become an industry framework, with YouTube and TikTok agreeing to the same restrictions so that tightening Instagram’s rules doesn’t simply send teens elsewhere. On glasses, he defends the recording light and says Meta needs to explain its safeguards more clearly as the devices go mainstream.

ICYMI: Sources ep. 1 with Sam Altman

Available everywhere you get podcasts.

Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.

After an unreleased OpenAI model recently escaped its sandbox and hacked Hugging Face, Sam Altman tells me why the company slowed parts of its frontier research and delayed a major reinforcement-learning run: “Getting AI safety right is more important than any company’s momentum.”

We also discuss Astra, OpenAI’s next-generation family of models, which is…

Watch or listen to the full conversation on Tuesday, September 1st, available everywhere you can get podcasts.

Yesterday, TIME and I published a cover story called “Inside OpenAI’s Reboot.” I spent two weeks at OpenAI’s San Francisco headquarters and spoke with more than 20 executives, investors, customers, and rivals.

I went into the reporting process wanting to understand OpenAI’s last year, what allowed Anthropic to gain so much ground, and the internal mood a…

Read more

My social feeds have recently been dominated by buzz for two new AI products: Elon Musk’s Grok Bot and an invite-only, text-based AI assistant called Instinct.

While the latter has yet to officially emerge from stealth, the rumor circulating is that Instinct’s small team, which includes an early Sierra researcher named Noah Shinn, is currently raising a Series A round of funding to buy more compute and open its waitlist. Like Poke before it, which was recently sold to Cognition, Instinct seems to be orchestrating off-the-shelf models with some specialized fine-tuning and tool use. It’s unusually capable and proactive at web-related tasks. There are a lot of little nice touches that help onboard a user more quickly, like the agent automatically researching your background while you’re setting it up. Early user takeaways often describe it as “OpenClaw for normal people.”

Read more

There are several ways to interpret the implications of OpenAI’s moves this week for Anthropic.

OpenAI’s largest frontier reinforcement-learning run remains on hold while the company studies what Sam Altman described to me as “various degrees of misalignment.” OpenAI says its expanded monitoring adds roughly 20% to the inference compute being monitored. That isn’t 20% of all internal inference, but it is a substantial compute tax. If the company made this move as a public stock, it’s hard not to imagine a strong market reaction.

Anthropic’s alignment work, meanwhile, could simply be ahead of OpenAI’s. Its internal frontier models could be less capable, meaning they aren’t showing the same problems yet. Or Anthropic could be seeing warning signs it hasn’t discussed, with a model jump forcing the issue soon.

Read more

“I think it is a good time to slow down,” OpenAI CEO Sam Altman told me last week, describing the company’s decision to halt the development of its most powerful, unreleased models.

The company announced Tuesday that it’s implementing new safeguards that will slow its future AI development. The company recently paused training on its next set of models, codenamed Astra, for a little more than two weeks, according to executives, and its largest planned frontier training run remains on hold while the new guardrails are put in place.

It is the first time OpenAI has made such a move. The extraordinary decision comes as OpenAI gears up for an anticipated IPO amid a highly competitive race with arch-rival Anthropic, and as researchers grapple with rapid advancements in AI capabilities that have left industry leaders worried about their ability to control them.

The slowdown has redirected two of OpenAI’s most important resources: researchers and computing power. Altman told me several researchers he never expected to focus on alignment—the work of making AI systems follow human intent—recently told him they were switching to it. “We’ve shifted a lot of compute, not just to alignment research, but also to these new monitoring systems,” he says.

Read more

It has been awhile since I’ve done a subscriber survey. If you have a moment, consider answering a few questions about yourself. Results are anonymized, and this data helps me build a better media business. It’s coming up on a year since I launched Sources, and I’m so grateful for all your support.

The trick to raising $5 billion, according to Ali Ghodsi, is not to plan on raising $5 billion.

Databricks was heads down preparing for its Data + AI Summit in June when The Information called to say it was about to report that the company was raising money at a $175 billion valuation. “We were just scratching our heads,” Ghodsi told me Thursday. Databricks asked the publication not to print the story. It did anyway.

Then Ghodsi’s phone blew up.

Read more

Over the weekend, Mark Zuckerberg called me through his Meta glasses while speeding across the water toward a fishing spot. The noise cancellation was holding up against what sounded like a wind tunnel. Before he got there, he wanted to walk me through the 6,500-word pro-AI manifesto he had spent the past month or two writing.

Read more