“Developers are the difference between a world full of slop and a world full of software worth using.” I loved this line from @rauchg in his opening keynote at @vercel Ship last week. It gets to the heart of why I don’t buy the doom and gloom for SWE jobs. The data backs it up - since Claude Code was introduced in late Feb ‘25, the number of job postings for software developers in the US has risen almost 15%. The agent writes the code, but the developer decides what's worth shipping.

It took 10 years to get to 1B @vercel deployments. Now, developers & agents ship 1B every 4 months. While the deployment pace is much, much faster now, @rauchg's mission hasn't wavered: create the fastest, most powerful infrastructure in the world and give that power straight to developers. Awesome to see a packed house in SF at Vercel Ship 2026. I'm especially excited to catch G's fireside chat with @tobi later today.

Who actually buys GPU compute? Very few companies, as it turns out. 9% of @tryramp customers paid a GPU vendor in the past year, against 78% that paid for an AI application. That small group spent nearly 80% as much as everyone buying the apps. How they buy is more telling. 8% bought compute through model serving and inference providers, 3% through neoclouds, <1% through wholesale capacity. Buyers would rather pay for finished work than rent a GPU and sort out the rest. That’s what we at @notablecap call workload clouds and it’s the bet @burkaygur and @gorkem made for generative media with @fal, what @rauchg made with @vercel, and what @pk_iv made with @browserbase.

Muse got to 5M downloads in 22 days, among the fastest apps to this milestone. For example, it took ChatGPT 56 days and Grok 103 days. My colleague @dan_cahana recently unpacked why personal agents like Muse and Instinct suddenly feel like they “just work,” becoming wildly popular with consumers. A big part of the answer is infra magic behind the scenes, including: -Model progress: LLMs have gotten dramatically better at complex, multi-turn task and tool use. -Cheap inference: Unlike chatbots, personal agents work in the background. Agent providers can trade off cost for latency, performing even more valuable tasks without incurring enormous token bills. Better models + better infrastructure = remarkable UX that wasn’t possible before. Great chart from @coatuemgt.

For decades, the basic unit of software consumption was the seat. Whether a user opens the application 2x or 20x mattered very little once the monthly bill arrived. Agents break that model. They work continuously, cross applications & execute thousands of actions without a person opening a tab. That raises one of the defining questions for agentic software... how does pricing change when software never stops working? A flat subscription gives customers predictability, but leaves vendors absorbing variable inference costs. Pure usage-based pricing protects margins, but makes enterprise spending harder to predict. Outcome-based pricing better aligns cost with value, but requires vendors to assume more execution risk. For infra companies, demand will increasingly be shaped by agent-hours, tool calls & the cost of maintaining persistent context, not just monthly active users. For enterprises, software procurement may begin to look more like workforce planning: -Which agents are authorized to act? -What applications can they access? -How much should each one be allowed to spend? -How is their performance measured? The agent is becoming a new unit of software consumption. Curious to hear from founders who are thinking about changing their pricing (or already have).

"Relatively overnight, no one really integrates the Stripe API by hand anymore. They send their coding agent to do it." I've been thinking about what this means for DevRel. Trust with developers is hard to earn and easy to squander. It's why in the early days at @HashiCorp, nearly half the DevRel team worked on nothing but content: documentation, tutorials, technical writing. The things a developer reaches for at 2 a.m. when something is broken. Now there's a debate about whether DevRel still matters once agents are choosing the tools. Before any human reads your docs, an agent has already tried to use your product. In my view, DevRel matters as much as ever. Documentation carries more weight than it ever has, and its reader is as likely to be an agent as a human. The best modern documentation caters to agents as much as humans. Awesome talk with @collision & @MollySOShea at @browserbase Navigate.

One of the most important questions for software companies today: how do you win the agent? As agents become users and increasingly make decisions about which tools and infrastructure to use, the playbook for product is changing. Quickly. I spoke to @alix_coutures about what @Vercel is seeing firsthand. With over 480,000 paying customers and an army of agents shipping code for each one, Vercel has a unique view into this shift. In June, @rauchg shared that agents that produce new code now account for half of Vercel’s new business. There’s a lot more to come as agents become customers. And Vercel is leading the way 🚀

Agents allow developers to break free from the constraint that one developer = one unit of labor. That fundamentally alters the math of the software industry. I sat down with @dan_cahana to unpack our recent essay on the developer as the bellwether for the AI era.

Amazing to be at the @fal Generative Media Conference today, where I got a chance to sit down with Sean Bailey, a true Hollywood legend. Sean co-founded LivePlanet with Ben Affleck and Matt Damon in 2000, before joining the Walt Disney Company as the President of Worldwide Production for 14 years. Our conversation about the intersection of AI and storytelling was timely after the @jeffreykWNDR essay published this week, which speaks to the collapsing divide between Hollywood and Silicon Valley. Sean spoke about how in Hollywood, the word “talent” means folks like the directors, the screenwriters, the actors. At Bay Area-based Pixar, talent meant every single person in the building, including the engineers. As the founder of B5 Studios, Sean is already seeing a shift, albeit a slow one, in the way the industry talks about AI. What used to be a clear binary (did you use AI or didn’t you?) is now being understood as hundreds of gradations of what AI can do. He sees the biggest opportunity now in upending the traditional linear creative process: pre-production, production, post-production. Animators can watch and iterate throughout the entire film creation process. That same flexibility simply didn’t exist in the live-action realm - there was only so much you could change on the fly or in post. The human element is still critical because the models will never be able to create the twists and emotional turns that pin us to our seats. But now, live action creators will also be able to iterate on the frontend and adjust on the backend, and the end product will be all the better for it. Disney pioneered the flywheel approach (a movie begets toys your kids adore, rides at the theme park, etc.), and Sean believes AI will create a new flywheel, where you can engage with the characters you love in formats we haven’t even conceived of yet. Thank you, Sean, for joining me, and to @burkaygur, @gorkem, @isidentical, and the entire fal team for putting on an amazing conference. It’s incredible to be back at GMC and see how much it’s grown in just its second year.

Developers have been the technology industry’s most reliable bellwether for 20 years. Their choices foreshadow the infrastructure shifts that follow. @dan_cahana & I wrote up what the bellwether is saying now.

Deployments on @vercel are up to 10M per day. Just a year ago, that number was 1M. Vercel CTO Malte Ubl (@cramforce) shared that eye-popping stat onstage at @browserbase Navigate last week. We’ve seen an explosion in software creation, and we’ve still barely scratched the surface of what agents will create. Every software company needs to figure out how to either get in the agent path or leverage the growth being created by agents.