Alexander Lovell

Most people still think of AI agents as chatbots with tools. That was the 2025 mental model. We’re already past it. The realization reshaping the architecture conversation: most agent failures aren’t intelligence failures. They’re failures of temporal coherence. Agents lose track of what they already tried, why prior strategies failed, what’s still uncertain. They have memory. They have almost no metacognition. The frontier is moving away from bigger prompts, more tools, and “AI employees” — toward persistent cognitive runtimes with bounded execution, verifier layers, trust-tagged memory, and runtime supervision instead of giant system prompts. The analogy that’s stuck with me is Erlang/BEAM. The architecture actually fits: lightweight isolated cognitive processes, message passing, supervisor trees, restart semantics. The agent failure modes I keep seeing in the wild are essentially the problems BEAM was designed to solve — not at the intelligence layer, but at the runtime layer. By end of 2026 I expect the first genuinely reliable long-running local-first agent runtimes. Coding, research synthesis, operations. Systems that hold coherent bounded execution over hours or days instead of minutes. That sounds less dramatic than the AGI narrative. It may turn out to be more important.

The wrong question is whether AI is overhyped. It is overhyped. So was e-commerce. So was big data. The useful question is which kind of overhyped technology AI turns out to be. Big data was real, but the hype ran well ahead of the value. During my time at Fivetran, I saw this pattern constantly: companies had more data than ever and still struggled to turn it into better decisions, faster operations, or durable advantage. They often confused the size of the data estate with the economic value of the data. Those are not the same thing. The value depended on plumbing, quality, access, ownership, workflow integration, and whether the organization actually changed how decisions got made. A lot of companies got pieces of that right. A much smaller number got the full value. E-commerce followed a different path. The 1990s hype was early and often ridiculous, but the long-term impact was larger than many skeptics expected. The world slowly reorganized around it: payments, logistics, search, advertising, marketplaces, inventory systems, and consumer behavior. The first-order version looked like stores with websites. The important version changed the operating model. AI could still become a big-data-like story: widely adopted, genuinely useful, but unevenly translated into enterprise value. That is the skeptical case I take seriously. My bet is that it looks more like e-commerce, and potentially much larger if recursive self-improvement becomes real. AI is not just a substrate to store, move, or analyze. It can act inside workflows. That means the ceiling is not only better reporting or better software. It is a different allocation of work, judgment, escalation, and accountability. The optimists are probably directionally right and temporally early. The skeptics are right to demand proof, but slow absorption is not the same thing as weak technology.

The most important aspect of a potential SpaceX IPO may have nothing to do with rockets. Public market investors may soon get a liquid vehicle combining launch, telecom, AI infrastructure, frontier-scale model training, and long-horizon compute deployment under a single capital structure. Historically, those businesses were financed very differently. Scientific and industrial compute infrastructure was treated more like utilities or heavy industry: high capex, long depreciation cycles, constrained margins, continuous reinvestment. Frontier AI companies have largely been priced like software platforms, with valuations tied to growth, capability leadership, and distribution advantages. SpaceX increasingly looks like an attempt to collapse those categories together. Viewed through that lens, some of the more unusual pieces start to fit. Starlink functions as a globally distributed network layer. Launch capacity becomes a supply-chain and deployment advantage for compute infrastructure. Even the orbital data center discussions make more sense. Whether this was centrally planned from the beginning or emerged opportunistically almost does not matter at this point. The pieces are converging anyway. AI-for-science workloads expose the infrastructure assumptions underneath this model quickly. Climate simulation, astronomy, protein folding, and large-scale robotics training all depend on sustained access to tightly coupled compute over multi-year periods. These workloads care less about benchmark optics and more about reliability, networking behavior, checkpoint integrity, and planning stability. That creates an interesting tension in the current market. Investors still appear willing to apply software-style valuation frameworks to businesses increasingly shaped by infrastructure economics: power procurement, accelerated hardware depreciation, networking constraints, cooling density, and massive ongoing capex requirements. Demand for AI compute will likely remain structurally high. But that does not necessarily mean all AI-linked infrastructure deserves software-era multiples. SpaceX may ultimately become one of the largest industrial compute buildouts ever financed under software-market assumptions.

One of the more important AI market shifts this year is that xAI increasingly looks less like a frontier lab and more like a capital-intensive infrastructure business trying to preserve frontier-lab multiples. You can see it in the recent Google and Anthropic compute deals tied to SpaceX/xAI capacity. The interesting part is not the revenue headline. It is that several people around large-scale training workloads believe frontier labs are internally discounting xAI-associated capacity relative to traditional hyperscaler environments. The concern is not the GPUs themselves. It is whether Colossus was built too quickly to inspire long-term confidence around reliability, redundancy, power architecture, cooling, and operational maintenance. That distinction matters at frontier scale. Training runs are sensitive to cluster instability, networking behavior, checkpoint reliability, and operational interruptions. Labs do not want surprises halfway through a multiweek run. Especially when they already have alternatives. I suspect this is part of why some of these arrangements appear structured more like flexible bridge capacity than deeply strategic long-term infrastructure commitments. Fine print matters. Some reports suggest at least portions of the agreements can be terminated on relatively short notice. This becomes especially important in the context of a potential SpaceX IPO. Depreciation is the core economic reality of the data center business. AI demand is enormous today, but these assets burn down fast and require constant reinvestment. If customers view the capacity as temporary overflow supply rather than trusted long-term infrastructure, the valuation framework starts looking very different from a pure frontier AI narrative. I am confident the demand is real. I am much less confident the market has settled on the right way to price the durability of that demand versus the depreciation curve underneath it.

Oh look, what we did at Google Cloud worked out pretty nicely. Take a bow, team https://lnkd.in/g6EN6tP2

Three years ago, my CEO interrupted my launch review at Fivetran. We had just shipped the biggest product of my tenure. An AI platform that compressed our connector ship time from six months to four weeks. A technical victory. I was walking the room through the metrics when he stopped me. “So… what do product managers do now?” I probably laughed. Probably said this was just the beginning. I rode BART home to San Francisco that afternoon. Drizzling. Fog low enough I couldn’t see the tops of the buildings. Somewhere on that walk, I realized I didn’t have an answer. Not for him. Not for myself. That question is why I left Fivetran. Yesterday in Tokyo, I gave the answer to 500 senior PMs at the Findy Product Management Summit. It took me three years to write. The talk was called The Agency Shift. Software is no longer a warehouse you go to and ask for things. It’s becoming a factory that produces outcomes. Sierra resolves the ticket. Harvey drafts the contract. Cursor writes the code. You don’t retrieve information anymore. You describe an outcome and the software produces it. This sounds abstract until you watch $285B vanish from SaaS valuations in 48 hours, the way it did in February. General-purpose coding agents collapsed the cost of custom software. Contracts and distribution are buying most companies 18-36 months. After that, what’s defensible when anyone can build the features? The argument I made is something I called the Memory Moat. Four layers a competitor can’t copy: data, process, decisions, compounding advantage. The next trillion-dollar companies will be built at layer four. Like Sierra, where every interaction across thousands of customers makes the system better for all of them. The deeper layers need three things AI can’t commoditize. Judgment about what your product is accountable for. Context that compounds with every customer. Trust infrastructure that earns the right to do more on the customer’s behalf. Judgment, context, trust. JCT. Product management was never about shipping features. It was about defining accountability, understanding the customer’s reality deeply enough to build for it, and earning the trust that lets the product grow. Features were the medium. JCT was always the work. The shift to agency didn’t change the job. It revealed it. The PMs who thought the work was features will find them built by agents in a weekend. The PMs who understood the work was always JCT will become more valuable, not less. I’m not certain I got every part right. I’ll know in 18 months. After carrying that question for three years, I’m grateful I got to put my best answer in front of a room of operators who could push back. They did. The questions were sharper than I expected. I’m already rewriting two slides. Thanks to Yuichiro Yamada, Masakazu Inaba and the Findy Inc team for hosting, and to Shinichi Kimura and Akira YAMADA for their coordination and conversation.

Thanks Findy Inc - this sure beats the soggy sandwiches I used to get as a speaker at Dreamforce or Next! Stay tuned for my keynote recording and interview next week on the The Agency Shift and what it means for the future of product managers

I’m getting awfully excited to give the keynote address at Findy Inc’s AI product management summit in Tokyo tomorrow. Register now to get the livestream! https://lnkd.in/gWqm8NHA

I spent a weekend trying to run local AI models on an Xbox One. The GPU seemed promising. It was just sitting there. How bad could it be? BAD, SO BAD. 8GB of 2013-era shared RAM, no Linux boot path worth debugging, zero inference acceleration. I gave up after an hour and googled “cheap mini PC.” Bought a $600 AMD box instead. That part took 15 minutes. Here’s my thing: every guide to building an AI dev environment is written by software engineers for software engineers. I’m a product executive. I want to build a blog, prototype an app, draft a board memo, and plan a family trip — ideally in the same afternoon, with the same system, without starting over every time I change jobs. So I built something different. Three principles drove everything: The system uses my real knowledge. Real email, real calendar, real notes (scary!). The system survives a job transition. Git-based knowledge, personal hardware, model-agnostic architecture. When an employer locks down comms, the personal layer persists. The system ships real things, not productivity workflows. The core: a $600 AMD mini PC running Ollama for local inference, Claude Code for building, OpenRouter to route between Claude, Gemini, and DeepSeek depending on task and budget. Privacy is where I made deliberate choices. Anything touching real files, contacts, or sensitive research runs through local models on my own hardware — never leaves the machine. Cloud models only handle tasks where quality matters more than privacy. For email and calendar, I use Claude Cowork’s SOC2-managed connectors rather than OpenClaw — fewer moving parts, no self-managed security surface, no CVE exposure I have to track myself. Total hardware: ~$650. Monthly API spend: $30–150 depending on how hard I’m building. What it doesn’t do well: local inference above 8B models is slow. Multi-agent workflows still break in interesting ways. Happy to share the full tech stack in the comments — hardware, model routing logic, and the one hardware detour I won’t mention here.