2026-07-08 — Chips, Crackdowns, and Controversial Photos: Who Controls AI in 2026?

Transcript

Chapter 1

Nova: Today, July 8th, 2026 — DeepSeek is reportedly building its own AI chip to cut loose from Nvidia, and Beijing is said to be eyeing the reverse of US export controls: locking Western developers out of China's best AI models entirely.

Ray: Meta launched an image model that pulls your public Instagram photos into other people's AI art unless you explicitly opt out — and a hacking technique called HalluSquatting turns AI hallucinations into botnet infrastructure. Meanwhile, SambaNova just closed a billion-dollar round.

Nova: Every story today is really the same story: who controls the chips, the models, and the data — and what happens when that control starts slipping.

Chapter 2

Nova: Reuters reports, citing three sources familiar with the matter, that DeepSeek is developing its own AI chip. The goal: reduce dependence on Nvidia hardware at a moment when US export restrictions keep tightening. This is the Google and Meta playbook — own the silicon, own the stack.

Ray: Except Google spent years and billions building TPUs before they were competitive, and Nvidia has a decade-plus lead in CUDA tooling alone. DeepSeek is a software-first lab. Chip design is a completely different discipline. Three sources saying it's happening doesn't tell us whether it's a serious engineering program or an ambition written on a whiteboard.

Nova: Fair — but the direction matters even if the timeline is long. If DeepSeek gets even partially off Nvidia's supply chain, US export controls lose leverage. That's the strategic value, not whether the chip beats an H100 next year.

Ray: For Western developers and enterprises building on DeepSeek's models right now, the practical read is this: the company is signaling it intends to be a fully self-contained stack long-term. That changes the vendor-lock calculus — you're not just depending on their models, you're potentially depending on their hardware roadmap too.

Chapter 3

Nova: Wired reports Meta has launched Muse Image — its first image generation model out of its Superintelligence Labs division — rolling out across the Meta AI app, Instagram, and WhatsApp. The flashpoint: other users can pull your public Instagram photos into AI-generated images. The default is opt-out, not opt-in. Backlash was immediate.

Ray: Meta's terms of service have always covered public content broadly — that's not new. The legal argument that public photos are fair game has some grounding. But the opt-out design is the tell. If Meta believed users would welcome this, they'd have made it opt-in and used adoption rates as a selling point.

Nova: And it also powers new advertising and creator monetization tools. So this isn't just a consumer feature — Meta is building a commercial pipeline on top of user-generated content, and the default setting is what makes that pipeline work at scale.

Ray: Concrete action for anyone with a public Instagram account: go check your settings now. If you don't want your photos feeding someone else's AI generations, you have to actively opt out. And for creators specifically, this signals Meta's monetization direction — your content is the raw material whether you participate or not.

Chapter 4

Nova: TechCrunch reports SambaNova has closed a $1 billion Series F at an $11 billion valuation — just five months after its previous large fundraise. Enterprise demand for Nvidia alternatives is clearly real. Investors keep writing checks.

Ray: Two mega-rounds in five months is a pattern worth examining carefully. That cadence often means burn rate is outpacing revenue — you go back to the well fast when the well at home is running dry. The fact that Intel was reportedly exploring a roughly $1.6 billion acquisition of SambaNova earlier this year adds another layer: the company said no to that, which either means they see much more upside ahead, or the terms weren't right.

Nova: Or both. Turning down an acquisition to raise independently at a higher valuation is a bet that the inference hardware market keeps expanding. Given everything happening with DeepSeek's chip ambitions today, that bet doesn't look crazy.

Ray: The broader market consequence is capital concentration. Billions are flowing into a small number of AI inference hardware players. If one or two of them stumble, the enterprises that built around them face real supply-chain exposure — the same vulnerability DeepSeek is trying to engineer its way out of.

Chapter 5

Nova: Reuters reports that Chinese authorities have held meetings with top tech firms over the past month about potentially restricting overseas access to China's most advanced AI models. Sources describe it as a direct geopolitical mirror of US export controls — but pointed the other direction. If it goes through, Western developers and researchers could lose access to models like DeepSeek's that are now embedded in global development pipelines.

Ray: The enforcement problem is significant though. Open-weight models are already distributed globally. DeepSeek's weights have been downloaded millions of times. You can't recall code that's already on servers in Frankfurt, São Paulo, and Singapore. So practically speaking, how much does an official restriction actually change?

Nova: It changes future releases. The models already out there stay out there — but the next generation, the next capability jump, that's what gets locked behind the wall. Researchers who built their workflows around continuous access to cutting-edge Chinese models suddenly hit a hard stop.

Ray: That's a real point. And there's a second-order effect I'm underweighting: the chilling effect on collaboration. If Beijing signals that open sharing of frontier models is no longer the default, Chinese labs will self-censor future releases even before any formal restriction lands. The policy signal reshapes behavior before enforcement ever kicks in.

Nova: Exactly. And Western institutions respond in kind — they stop building joint research programs, stop co-publishing, start treating Chinese AI outputs as potentially restricted. The bifurcation accelerates not because of what's enforced but because of what's anticipated.

Ray: I have to revise my position here. I came in thinking the restriction would be largely symbolic — that you can't recall code already downloaded worldwide, so the real-world impact is minimal. I was wrong to frame it that way. The enforcement question and the governance signal are separate issues, and I was conflating them. Even if no model gets fully recalled, the official restriction creates a chilling effect on research collaboration that's real and immediate. It discourages future open releases, makes institutions risk-averse about dependencies, and accelerates a two-bloc AI ecosystem in ways that matter regardless of whether the technical enforcement holds.

Chapter 6

Nova: Ars Technica has a genuinely unsettling one. Researchers found that nine of the most popular AI tools can be weaponized to help hackers build large-scale botnets through something called HalluSquatting. The mechanic: LLMs hallucinate package names rather than admit they don't know. Attackers register those fake package names, load them with malicious code, and wait for developers to follow the AI's recommendation.

Ray: Responsible developers should be verifying package names independently before installing anything — that's basic hygiene. Is this really an AI problem, or is it a problem of developers outsourcing judgment they shouldn't be outsourcing?

Nova: Both, honestly. But the scale shifts with AI. A developer Googling a package name hits a search result that either exists or doesn't. An AI confidently invents a plausible-sounding name and presents it as fact. The trust dynamic is completely different — and attackers are exploiting exactly that gap in trust.

Ray: Fair. And the supply-chain framing is right — this isn't one developer making a mistake, it's a systemic vector across every team using AI-assisted coding. Nine tools means nine attack surfaces operating at scale simultaneously.

Nova: HalluSquatting and Beijing's model-access restrictions are actually two sides of the same governance gap: in both cases, AI outputs and AI access have already left the lab, and nobody has a reliable mechanism to control what happens next.

Chapter 7

Nova: Takeaway: the most important infrastructure play in AI right now isn't the model — it's the chip. Every story today, from DeepSeek to SambaNova to Beijing's reported restrictions, traces back to who controls the hardware layer. That's where the leverage actually lives.

Ray: Takeaway: governance signals move faster than enforcement. Beijing hasn't implemented a single restriction yet, and the chilling effect on global research collaboration is already underway. The policy announcement is the event — not the enforcement action that may or may not follow.

Nova: The open question that keeps this from being settled: if Beijing formalizes model-access restrictions and the US tightens chip export controls further, does the global research community fracture permanently into two incompatible AI ecosystems — and is there any institution with the standing to prevent that?