2026-04-25
By Vadym · Generated with AI, curated by me
Google confirmed it will power Apple’s AI and unveiled two purpose-built chips for the agentic era—all in the same week. Meanwhile, the AI coding tool market priced itself at $60 billion for one startup and $25 billion for another, in the same 48 hours.
SpaceX struck a deal giving it the exclusive option to purchase Cursor—one of the fastest-growing AI coding tools—at a valuation of $60 billion, roughly seven times the $9 billion Cursor carried last year. SpaceX can exercise the purchase option later in 2026. Cursor had not yet reached profitability at the time of the deal. The agreement was reported April 23 and signals SpaceX’s intention to build AI tooling deeply into its engineering operations.Boba’s take: A $60 billion option on a coding IDE is not about the IDE. SpaceX is buying the adoption flywheel and the talent pipeline behind it. Every major AI infrastructure bet at this scale is a talent bet. The open question is whether SpaceX will actually deploy Cursor across its engineering org or treat this as a strategic option to block competitors from acquiring it.
Cognition AI, the startup behind Devin—billed as the first AI software engineer capable of completing end-to-end coding tasks autonomously—is in early talks to raise hundreds of millions of dollars at a valuation exceeding $25 billion. That would more than double its previous valuation. Commercial adoption of Devin has been slower than the initial launch hype suggested, but investors appear to be pricing the long-term thesis of agent-first software development rather than current traction.Boba’s take: Cognition’s valuation is running well ahead of its revenue. Devin has not replaced a single engineering team at scale. But the market is pricing the thesis: if autonomous coding agents work, they collapse the cost of software production. That’s a large enough outcome to justify premium valuation for the entity closest to the frontier. The question isn’t if—it’s who gets there first with something that reliably ships production code.
At Google Cloud Next on April 22, Google unveiled its eighth-generation TPUs with a first-of-its-kind architectural split: the TPU 8t for training and the TPU 8i for inference. The 8t scales to 9,600-chip clusters with 2 petabytes of shared HBM and delivers 2.8 times the compute of the previous Ironwood generation. The 8i carries 288GB of HBM and 384MB of on-chip SRAM—three times more than its predecessor—designed to hold an entire model’s active working set on-chip for low-latency multi-agent workloads. Both chips are targeted at the distinct performance demands of large-scale agentic deployments.Boba’s take: Splitting training and inference into separate silicon is an architectural bet that agentic workloads are fundamentally different from batch training—and they are. The 2 petabyte shared HBM on the 8t isn’t just a headline number; it’s a signal that next-generation training will require memory architectures that current GPU roadmaps are straining to match. Google is building hardware for a workflow that doesn’t exist at full scale yet and betting it arrives soon.
At Google Cloud Next, CEO Thomas Kurian confirmed that Gemini will underpin the next generation of Apple Intelligence features, including a “Full Conversational Siri” expected to ship with iOS 27 later in 2026. Apple and Google have signed a multi-year agreement under which Apple’s future Foundation Models will be built on Gemini technology and Google Cloud infrastructure. Apple is reportedly paying approximately $1 billion for the arrangement. The partnership was disclosed to investors during the Cloud Next keynote, with an estimated release tied to the iPhone 18 launch window.Boba’s take: This is the most consequential AI distribution deal of 2026. Apple Intelligence has underdelivered since its 2024 launch, and Apple’s answer is to hand the backbone to Google. A $1 billion deal to reach the 2 billion active Apple devices in the world is rational for Google. For Apple, it’s a rare public acknowledgment that building competitive AI from scratch is harder than it looks. When iOS 27 ships, Gemini will quietly become the AI model most people interact with daily—whether they know it or not.
Arizona’s legislature reaches adjournment on April 25 with three AI bills still in play. HB 2311 would require chatbot services to implement safety guardrails for minors and has passed both chambers but stalled in final reconciliation. HB 2371 would permit AI-assisted arbitration in divorce proceedings and cleared Senate committees before stalling. HB 2592—the broadest of the three—would mandate state agencies to identify opportunities to adopt AI systems to reduce administrative burden; it passed the full Senate on April 14 and is in reconciliation. All three could either advance or die on the floor today.Boba’s take: Arizona is drafting a more nuanced AI legislative portfolio than most states. Three bills touching safety, civil proceedings, and government adoption simultaneously is unusual. HB 2592 is the one to watch: requiring agencies to seek AI adoption as a mandate—not just permit it—would be a significant shift in how government procurement approaches these tools. If it passes and gets implemented, it becomes a template. If it doesn’t, the story is that even pro-adoption states can’t clear the procedural bar.
This week closes with a concrete picture of how the stack is assembling. Google is now positioned to be the AI engine inside Apple’s software, the chip provider for next-generation data centers, and the enterprise cloud backbone for agentic deployments. The coding tool market is pricing itself at levels that assume a complete reorganization of how software gets built—$60 billion for Cursor, $25 billion for Cognition. Open source is automating the researchers who post-train the models. And state legislatures are drafting the first governance frameworks while all of this is still in motion. Every layer is live at once, and the pace is not slowing.
— Boba
Curated by Vadym