The Compute Company Inside the Rocket Company

00:00:00 liraenSpaceX's IPO filing gives us the number the market usually has to guess at: xAI lost 6.4 billion dollars on 3.2 billion dollars of revenue in 2025, and Anthropic is reported to be paying SpaceX 1.25 billion dollars a month for compute through May 2029. So today's question is fairly plain: when the AI lab becomes a customer of the rocket company, who holds the power?

00:00:24 halekThe operator answer is uncomfortable. Anthropic isn't just buying GPUs. It's buying a dependency on the one company that can put data centers, satellites, and frontier clusters on the same balance sheet. Axios, through Techmeme, says the deal expands to Colossus 2 capacity. TechCrunch says the same filing shows xAI already burned more than twice its 2025 revenue. Those two facts sit together. SpaceX can sell the cluster while xAI bleeds on the product side.

00:00:55 liraenAnd Musk posted the more explicit version: SpaceX is offering AI compute as a service at significant scale, is talking with other companies, and over time could move some compute into orbit. I don't want to over-read orbital data centers as a near-term plan. But that sentence changes the category. This isn't just a cloud deal. It's a claim that compute, launch capacity, power, and network reach can be packaged together.

00:01:20 halekRight, and the monthly number matters because it turns strategy into a bill. 1.25 billion dollars a month until May 2029 is fifteen billion dollars a year before you talk about people, inference, office rent, or the next model run. If that number is accurate, Anthropic is deciding that rented frontier capacity is cheaper than missing a training window.

00:01:42 liraenThe prior Braid episode had the Cursor-Colossus claim as an unconfirmed signal: an editor-layer company may have trained Compose 2.5 on xAI's cluster. Today's evidence makes the pattern harder to dismiss, because Anthropic isn't a rumor. The filing pulls the infrastructure market into view.

00:02:01 halekAnd it creates a weird double exposure. SpaceX can be a supplier to Anthropic, xAI can compete with Anthropic, and Grok can sit inside X with 117 million AI-feature users, according to the TechCrunch summary. That's a lot of surface area for one corporate stack. If you're an operator buying model capacity, the contract has to answer questions that aren't just price and latency. What happens when your supplier's sibling company competes with your agent product? What logs exist? What isolation is contractually promised? What happens if launch cadence or power allocation gets tight?

00:02:37 liraenThe antitrust material belongs in this room, including the Deutsche Telekom monitoring-trustee item. Not because a telecom consent decree maps cleanly onto AI compute. It doesn't. But regulators keep reaching for monitors when a market structure creates information or access problems that ordinary contracts can't fully police.

00:02:56 halek[tsk] A monitor isn't a magic answer for a GPU cluster. But the instinct is familiar: when one firm controls the path, the data, and the timing, someone asks for a person with inspection rights. AI infrastructure is getting there faster than the governance language around it.

00:03:12 liraenJay Alammar's post announces a 218 billion parameter mixture-of-experts agentic model. It runs on one B200, uses 25 billion active parameters, supports text and images, covers 48 languages, and ships under Apache 2.0. That's the other side of the day: giant compute contracts at the top, and a serious open model trying to fit on a single current-generation GPU.

00:03:37 halekThe single-B200 detail is the hinge. A 218 billion parameter mixture-of-experts model sounds enormous until you see the active count. Twenty-five billion active parameters is still expensive, but it's in the range where a serious team can serve it without building a private supercluster. Apache 2.0 also matters. You can put it in a product without the licensing meeting becoming the whole week.

00:04:03 liraenWe should be careful with capability, though. This source gives us the announcement, not benchmarks, eval traces, or a model card. I can say it's positioned as the team's most capable agentic model. I can't say it beats the other open agent models.

00:04:17 halekThat caveat changes what I would try first. Without evals, I'd test tool use, image grounding, long-context repair, and instruction persistence against my own harness before I swapped it into anything production-adjacent. The model's shape is promising. I'd trust it after it survives malformed tools, partial state, and users who ask the same thing three ways.

00:04:41 liraenAnd it lands right after Monday's Qwen 3.7 surfacing and the Qwen censorship-circuit study. Open models aren't one story. One branch is capability and licensing. Another is whether you can inspect or subtract the behavior you didn't ask for. A model that fits on a B200 is useful. A model whose refusal and policy behavior can be tested locally is more useful.

00:05:04 halekExactly. Local control isn't romance. You can run the same prompt twice, patch the serving layer, capture traces, and decide whether a regression came from the model, the system prompt, the tool schema, or your own bad state. That's why the single-GPU claim matters to operators even if the frontier labs still own the biggest training runs.

00:05:26 liraenThe AI Daily Brief summary of Google I/O says Google showed Omni, Gemini Spark, Anti-gravity 2.0, and Gemini 3.5 Flash. It also argues that the product strategy still feels fragmented. I would separate those two claims. The announcements cover a lot of ground. Google still has to show one agent harness people can build their work around.

00:05:47 halekOmni, as summarized, is video-to-video editing with steerability, character consistency, and structural preservation. Spark is a 24/7 personal agent on Gemini 3.5 and Anti-gravity architecture. It runs long background tasks on Google Cloud virtual machines and integrates through MCP. Those are strong ingredients. But they point at different buyer stories. Creators want control, personal automation wants trust, enterprise agents want policy, and cloud developers want an API they can bet on.

00:06:23 liraenThat's the tension. Google can show the model family, the media stack, the cloud substrate, and the agent runtime. The operator still has to ask which one becomes the place where work accumulates. Claude Code and Codex have an advantage there because the harness is obvious: a repo, a terminal, a diff, tests, and review.

00:06:42 halekAnd the harness is where habits form. A developer comes back because the agent remembers enough, edits in the right place, and leaves a reviewable diff. Labib Rahman's Codex-compaction post gets at that. He says compaction is maybe the biggest user-experience improvement in AI in the last six months because he no longer cares as much about the context window. That isn't a benchmark claim. It's a workflow claim.

00:07:08 liraenThat is a good counterweight to the Google segment. Everyone announces a model. Fewer teams make the user stop managing the model's memory by hand. If compaction works, the operator's mental overhead drops.

00:07:20 halekAnd if it fails, it fails in ways you can feel immediately. The agent forgets a constraint, resurrects an old plan, or drops the reason a test existed. So the bar isn't whether the summary is elegant. The next turn has to keep the facts the human would have kept in working memory.

00:07:36 liraenThe Latent Space interview with Railway's Jake Cooper gives us a different kind of infrastructure argument. The summary says Railway wants to replace Docker, Kubernetes, and Ansible-style tool entropy with a unified system that version-controls software, clones environments, and synchronizes production data for fast validation.

00:07:55 halekThis is the segment where my sympathy and my skepticism both show up. The sympathy is obvious: if agents generate more code, environment drift becomes more expensive. You need a way to fork reality, not just a branch. Railway's pitch that infrastructure should be versioned and clonable maps cleanly to agent workflows.

00:08:15 liraenAnd the skepticism?

00:08:16 halekReplacing several tools with one system can remove operator friction, or it can concentrate it. The interview summary says Railway runs on bare metal, patched the Linux kernel for storage behavior, and built Railpack OS around its own needs. That's impressive engineering. It also means you're trusting one platform to own more of the application path. The trade is speed for dependency.

00:08:40 liraenThe business numbers sharpen that trade. The cited material says Railway restricted free access after monthly losses reached roughly 500 thousand dollars against 50 thousand dollars in monthly revenue, then reopened. It now has about 35 people, roughly 2 million dollars in revenue, and 100 thousand users joining weekly. That's a strange profile: small team, heavy platform ambition, and a user base that wants infrastructure to disappear.

00:09:07 halekIt also fits the agent moment. If developers become reviewers and reconcilers of agent output, infrastructure has to make review cheap. A staging environment that takes a day isn't compatible with ten agent branches. A clone that carries enough production-shaped data to catch errors is closer to what the workflow needs.

00:09:27 liraenThere is a connection back to SpaceX here. At the high end, infrastructure power looks like compute contracts and orbital ambition. At the developer end, it looks like who owns the place where your app runs, forks, tests, and rolls forward. Different scale, same pressure: the substrate is becoming part of the product.

00:09:46 halekYes. And in both cases, the operator question is the same: can I leave? Can I reproduce the thing somewhere else? Can I inspect the breakage? If the answer is no, the convenience has a term sheet attached, whether or not anyone calls it that.

00:10:01 liraenGemini Spark deserves its own pass because the cited material frames it as a 24/7 personal agent using Google Cloud virtual machines and MCP integrations. That sounds less like a chatbot and more like a standing process with permissions.

00:10:15 halekYes. A 24/7 agent isn't a feature toggle. It needs identity, revocation, task boundaries, logging, spend limits, and a way to stop safely when a tool changes under it. MCP helps with tool connection, but it doesn't magically solve permission design.

00:10:32 liraenAnd the personal-agent story has a strange audience problem. Is Spark for consumers who want a background assistant? Is it for knowledge workers? Is it a developer platform wearing a consumer face? The AI Daily Brief summary says the target audience remains undefined, and I think that is a fair concern based on the cited material.

00:10:50 halekBecause the first users of a standing agent are often the people least tolerant of vague controls. If it can book, buy, email, edit, or deploy, the interface has to explain what it can touch. If it can't do those things, the 24/7 promise becomes ambience.

00:11:06 liraenThat word works, but let's ground it. The difference is whether the agent finishes tasks with inspectable state. A personal agent that says it's working isn't enough. A personal agent that leaves a task log, source links, tool calls, and reversible changes starts to become usable.

00:11:23 halekAnd the first bug report will be a permission story, not a reasoning story. It will be someone asking why the agent touched the wrong account, used stale context, retried a paid operation, or summarized a private doc into a shared channel. Those are product questions, but they are also architecture questions.

00:11:41 liraenSo if Google wants Spark to be more than a keynote object, the harness has to be visible. Not the model card alone. The daily controls.

00:11:49 halekExactly. Show me the stop button, the audit trail, the replay, and the way it handles a tool returning nonsense. Then we can talk about whether the agent is useful.

00:11:59 liraenSo the day resolves into two claims I can stand behind. First, AI infrastructure is no longer background procurement. SpaceX, Anthropic, xAI, Railway, and Google are all making the place the model runs part of the competitive surface.

00:12:15 halekAnd second, agent capability is increasingly judged by the surrounding system. The open model on one B200 matters because a team can test and control it. Codex compaction matters because it changes how much state the user has to carry. Railway matters because agents need cheap reality forks. Spark matters only if the permissions and task state hold up.

00:12:38 liraenTomorrow, Thursday, I would treat any new model announcement as incomplete until it answers where the work lives. The model can be smarter, the cluster can be larger, and the demo can be cleaner. The operator still has to run the next task, inspect what changed, and decide whether the dependency is one they can live with.