2026-06-25 — Silicon, Spies, and Super PACs: Who Controls AI in 2026?

Transcript

Chapter 1

Nova: Today, June 25th, 2026 — OpenAI names its first custom AI chip after a pepper, Anthropic accuses Alibaba of illicitly accessing Claude at massive scale, and Google DeepMind ships autonomous computer use inside a budget model.

Ray: Qualcomm drops nearly four billion dollars on compiler software, and pro-AI super PACs burn through twenty-seven million dollars in a single New York House primary — and still can't buy a seat.

Nova: Chips, alleged espionage, agent wars, and ballot-box proxy fights — all on the same day. The question running through all of it: does anyone actually control AI, or is everyone just scrambling for the levers?

Chapter 2

Nova: TechCrunch reports that OpenAI has unveiled its first custom ASIC chip, named Jalapeño, built in partnership with Broadcom. It's designed specifically for AI inference — powering current and future large language models. OpenAI joins Google and Amazon as companies with proprietary AI silicon. The strategic read is clear: owning the chip means lower inference costs and real leverage against Nvidia's pricing power.

Ray: First-gen anything in chip design is a proof of concept, not a competitive weapon. Google's TPU program is over a decade old. Amazon's Trainium and Inferentia lines have gone through multiple iterations. Jalapeño is OpenAI's first attempt. The gap between a debut ASIC and production-grade silicon that actually displaces Nvidia at scale is enormous — and measured in years, not quarters.

Nova: The timeline criticism is fair, but the direction matters. Every inference cycle OpenAI runs on its own silicon is a dollar not paid to a third-party supplier. For enterprises building on OpenAI's API, this eventually translates to pricing stability — potentially lower costs — because OpenAI controls more of its own cost structure.

Ray: Or it translates to a massive capital sink that distracts from model research while Google and Nvidia continue iterating. Broadcom is a capable partner, but this is still a partnership — OpenAI doesn't own the fab, the process node, or the full supply chain. The dependence just shifted, it didn't disappear.

Chapter 3

Ray: The Google DeepMind Blog announced computer use capabilities inside Gemini 3.5 Flash — meaning the model can now interact with software interfaces and perform tasks autonomously on a computer. The fact that this lands in Flash, their fast and cost-efficient tier, is the tell. Google isn't positioning this as a premium feature. They're treating autonomous computer interaction as a baseline developer tool.

Nova: That's the signal worth paying attention to. Anthropic's Claude computer use was a differentiator when it launched. Google embedding the same capability in a cheap, high-throughput model means agentic AI stops being a premium moat and becomes a commodity. Developers building automation pipelines now have a cost-efficient alternative, and competition at that layer is good for the ecosystem.

Ray: Speed and low cost in an autonomous model also means fast, cheap mistakes at scale. A model that can interact with any software interface is also a model that can be manipulated through that interface — prompt injection through a webpage, a malicious UI element, a spoofed dialog box. Enterprises aren't ready for the liability surface that opens up when an AI agent can click, type, and submit on their behalf.

Nova: The liability question is real, but it doesn't belong only to Google. Every lab shipping computer use features faces it. The difference is that Google's distribution reach means these questions get stress-tested at a scale that will actually force the industry to develop answers faster.

Chapter 4

Nova: Wired reports that Qualcomm has agreed to acquire Modular — an AI chip software startup known for next-generation compiler and hardware abstraction technology — for close to four billion dollars. That's one of the largest AI-adjacent acquisitions of the year. The thesis is straightforward: chips alone aren't enough. The software layer that tells hardware how to run AI workloads efficiently is now the decisive battleground.

Ray: Four billion dollars for compiler tooling is an extraordinary bet, and the target of that bet is Nvidia's CUDA ecosystem — which has a decade-plus head start and developer loyalty that goes bone-deep. Qualcomm's hardware is capable, but the reason enterprises and researchers keep choosing Nvidia isn't just the silicon. It's the toolchain, the libraries, the institutional knowledge baked into millions of lines of existing code.

Nova: Modular's approach was specifically designed to abstract across hardware targets — the goal was always to make code portable, not locked to one vendor. If Qualcomm can deploy that abstraction layer broadly, they don't need to beat CUDA on its own terms. They need to make switching costs low enough that developers stop treating Nvidia as the only option.

Ray: That's the pitch, and it's a good pitch. But Qualcomm now has to execute it while integrating an acquisition, keeping Modular's engineering talent, and convincing a developer community that has been burned before by 'CUDA alternatives.' The software moat Nvidia holds isn't just technical — it's social. That part doesn't yield to a four-billion-dollar check.

Chapter 5

Nova: Bloomberg reports that Anthropic has accused Chinese tech giant Alibaba of orchestrating a large-scale effort to illicitly access its Claude AI models. The accusation raises serious concerns about IP theft and misuse of frontier AI systems. Bloomberg frames this as a landmark legal and geopolitical moment — a new flashpoint in US-China AI tensions that could prompt tighter access controls across the industry.

Ray: These are allegations. Bloomberg's reporting is serious, but Anthropic's accusation is not an established legal finding. The history of US-China tech disputes includes cases where accusations were politically timed, overstated, or later contested. The industry should not let geopolitical framing harden into policy before the facts are established through a legal process.

Nova: The legal status of the claim is one question. The underlying vulnerability it exposes is a separate one. If the alleged scale of access described by Bloomberg is even partially accurate, it reveals something Anthropic and every other AI lab has quietly known: API access controls were built for growth, not for security. Usage monitoring at the frontier model level has been an afterthought.

Ray: And the response to that monitoring gap, if it comes in the form of access restrictions tied to national origin, creates a different problem. Tighter controls that effectively exclude international researchers — including legitimate academic and commercial users — risk fragmenting the global AI ecosystem. That fragmentation serves protectionist politics as much as it serves genuine security.

Nova: The fragmentation concern is real. But the alternative — leaving frontier models accessible through controls that allegedly couldn't detect large-scale illicit use — is not a neutral status quo. It's a systemic gap. Anthropic's response, if it forces the industry toward stricter verification and usage auditing, produces an outcome the field actually needs regardless of how this specific case resolves.

Ray: I came into this holding that unverified allegations shouldn't drive policy and that access restrictions risk fracturing the global research ecosystem in ways that serve protectionism more than security. I'm shifting on that. The alleged scale of access Bloomberg describes isn't a one-off API abuse — it points to a monitoring infrastructure that was never built to catch this kind of activity. That's a systemic gap I can't dismiss as geopolitical noise. If Anthropic's response leads to real usage auditing across the industry, I now think that outcome is worth the geopolitical friction — even knowing those same tools could eventually be misused for exclusion.

Chapter 6

Nova: The Verge covered a story that feels almost satirical but isn't: a twenty-seven million dollar political battle in New York's 12th Congressional District, fought between pro-AI super PACs backed by interests aligned with Anthropic and OpenAI. The candidate they were backing, Assemblyman Alex Bores, narrowly lost his Democratic primary. The race became a proxy war over AI policy influence in Congress — and the massive outside spending failed to deliver a decisive outcome.

Ray: Twenty-seven million dollars in a single House primary, and the result was a narrow loss. There's something almost clarifying about that. Voters in New York's 12th District weren't moved to deliver a clean mandate by the largest AI-aligned political spend anyone has seen in a congressional race. That's actually a healthy signal — it suggests AI policy preferences aren't something you can simply purchase at the ballot box.

Nova: The outcome matters less than the precedent. AI-aligned super PACs just demonstrated they're willing to deploy this kind of capital in a single down-ballot primary. Win or lose, that changes the calculation for every congressional candidate thinking about AI regulation. The message to future legislators is: take a position on AI policy, and this money will show up in your race.

Ray: Which brings it back to the episode's central thread. Chips, model access, compiler software, legal accusations — all of those are fights over who controls AI infrastructure. This is the same fight, one layer up. If AI companies can shape which legislators write the rules, the governance question isn't being settled in regulatory comment periods anymore. It's being settled in primaries.

Chapter 7

Nova: Nova's takeaway: control over AI is fragmenting into every layer simultaneously — silicon, software, legal standing, and now electoral politics. Any company or government that focuses on only one of those layers is going to find the others have moved while they weren't watching.

Ray: Ray's: the Anthropic-Alibaba situation is the sharpest version of today's real question. If frontier AI labs built their access infrastructure for growth rather than accountability, who is actually using these models right now — and does anyone have a complete answer to that?