Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, I'm your boy Tony DeLuca, and today's menu has one of those stories that reads like a spy thriller but lands right in the middle of the AI chip business. Pour the coffee, settle in, let's have at it.
Let me start with the one that made me put the mug down and read it twice. There's a report out of the New York Times, which Ars Technica wrote up yesterday, October 29th, and the headline is a mouthful but it's worth every syllable. The claim is that President Trump was briefed that China had planted a spy close to Nvidia CEO Jensen Huang, and that he chose to do nothing about it, at least partly to avoid hurting Nvidia's stock. Let me say that again, because it's the kind of sentence you want to hear slowly. A sitting president, told by his own intelligence agencies that a foreign adversary had a line into one of the most important executives in the country, and the decision was to sit on it.
Here's how the reporting lays it out. About a year ago, according to people briefed on classified intelligence, Trump was told that Beijing was getting reports from someone with direct access to Huang. That someone is a woman named Alice Huang, also known as Huang Le. She's a Chinese citizen, a former top executive at a Singapore cloud company called Megaspeed, and to be clear, she is not related to Jensen Huang. The US already suspected she was helping China evade export controls, getting access to high-end Nvidia chips through Megaspeed. But what alarmed the investigators was the discovery that she may have used her role to get personally close to Jensen Huang himself, and possibly to influence how Trump was thinking about AI policy.
And there's a detail in here that tells you everything about how this relationship worked. There's a photo from an event where she reportedly bragged, and I'll quote it, "I bet you guys I can get Jensen here." And then he arrived, trademark leather jacket, took whiskey shots. Now, Nvidia's spokesperson, John Rizzo, pushes back hard on the idea that any of this is sinister. He says it's common for the CEO to attend parties with customers and partners, that Jensen talks to thousands of customer executives every year, and that these interactions were "not unusual or unexpected." A source told the Times that Huang had always planned to attend that particular party. So Nvidia's line is: this is just a CEO doing CEO things.
But here's the part that should stick with you as a builder, because this is where the espionage angle turns into a policy angle. The report says Trump was so aligned with Huang, who has become the president's key AI advisor, that action against the alleged spying could hobble Nvidia's chip sales or disrupt the delicate US-China trade truce. Treasury Secretary Scott Bessent said last month, ahead of Trump's AI talks with Xi Jinping, that "the president is completely aligned with Jensen Huang, the CEO of Nvidia." And the sources told the Times that some investigations and proposed trade sanctions involving China or Nvidia have, quote, "ground to a halt within the Commerce Department" because of Huang's influence.
Now let me sit on that for a second, because this is the thing I keep coming back to. We've spent the last couple weeks on this show talking about Anthropic's S-1, the risk factors, the existential warnings, the enterprise spend tools, all this stuff about how AI companies see their own future. This is a different flavor of risk entirely. This is the question of who's steering the ship on export policy, and whether that steering is happening because of national interest or because of one company's stock price and one executive's relationships. Since cozying up to Trump, the Nvidia CEO has gotten export controls relaxed and struck a deal giving the US a 25 percent cut of some chip sales so he could sell more chips in China. That's a real policy outcome. That's not a party photo. That's billions of dollars and the direction of American chip strategy.
And here's the piece that makes it genuinely murky instead of a clean scandal. Nvidia itself investigated Megaspeed. Their probe concluded there were no signs of diversion, that Megaspeed raised no red flags as a customer. Jensen Huang, according to what Trump was told, had no knowledge of the alleged intelligence-gathering at the time of the briefing. And there's a reason he may never have gotten a defensive briefing, the kind Americans sometimes get when they're targeted by foreign spies. US agencies may have worried he'd tip her off that she was being watched. So you've got a situation where the CEO might be genuinely in the dark, or, as some senior officials apparently believe, it's unlikely he'd have remained unaware for the entire duration of a professional relationship. The report doesn't resolve that. It can't.
But let me walk you through the Megaspeed red flags, because this is where it gets concrete for anybody who thinks about supply chains and compliance. The US government has been investigating Megaspeed since 2024. They found suspicious boxes of Nvidia equipment that looked ready to ship out rather than installed. Earlier this year, the Department of Justice took over the Commerce Department's investigation. The company insists it has Singapore-based owners, but the Times visited the supposed offices and found staffers who couldn't clearly say what Megaspeed does, and confirmed that the Singapore director, a man named James Tan, lives in China and is never present. Corporate records show Tan owns the shell company that owns Megaspeed. And Megaspeed arranged a two-billion-dollar purchase of chips over the next year, with concerns that those chips could end up serving customers in China through data centers in Malaysia and Indonesia that appear to remotely serve China.
Aaron Bartnick, who worked on technology security and governance under the Biden administration, told the Times that Megaspeed raised "many red flags," including ownership claims and possible ties to sanctioned entities, the kind of thing that should have made Nvidia's executives question whether this company should have had access to its chips at all. His line: "It's hard to see a justification for taking that risk." And the Chinese Embassy spokesperson, Liu Chang, called the whole thing baseless, accused the US of "groundless smears and slanders."
So where does that leave you, the builder, the founder, the person trying to figure out what any of this means for your roadmap? Here's my read. If you are building anything that touches the AI chip supply chain, anything that depends on export control stability, you now have to price in a world where the enforcement of those controls is subject to the mood of one CEO's relationship with the White House. That's not a conspiracy theory, that's what the reporting describes, investigations grinding to a halt at Commerce. When policy enforcement becomes personal and discretionary, your ability to plan around it goes out the window. The rules on paper matter less than who's in the room. And that's a genuinely uncomfortable thing to build a company on top of.
Now, I want to be careful here. This is a report built on anonymous sources describing classified intelligence. Nvidia denies the trade-secrets angle. Megaspeed denies wrongdoing. Nobody has been charged. So I'm not going to tell you a spy brought down American AI policy. What I'll tell you is that the questions this raises, about whose interests are shaping chip export decisions, are the kind of questions that outlive any single news cycle. Keep an eye on whether the DOJ investigation into Megaspeed actually produces anything, and whether those stalled Commerce investigations get un-stalled. That's the tell.
Alright, let me shift from the spy who came in from the data center to something that's going to hit your manufacturing costs directly. Staying in the land of Trump, China, and hardware, there's a second policy move worth knowing about. On October 28th, Trump issued a presidential proclamation declaring imports of foreign-made factory robots a national security threat. Now, I want to frame this correctly, because the FCC's broader robot restrictions, the ones that hit robot vacuums and robot dogs and humanoids, that was back in July. This new proclamation is the industrial-arm piece that the July action had specifically exempted.
Here's what's actually new and what matters. The proclamation orders negotiations with trading partners and holds out the threat of tariffs on industrial robots if no agreements get struck within 180 days. These are the big heavy robotic arms that run car factories and warehouses. And here's the kicker, the thing that makes this almost comic: the US doesn't make them. The "big four" industrial robot makers, FANUC, ABB, Yaskawa, and KUKA, are headquartered in Japan, Switzerland, and Germany. No US company offers comparable heavy-duty hardware. Trump's own proclamation admits this, that US robotics companies don't produce enough to meet national security needs, and that the country relies on foreign imports for most types of robotics and their components.
So think about who gets squeezed here. General Motors, Ford, Tesla, Boeing, Lockheed Martin, all of them buy these arms. The International Federation of Robotics counted nearly 38,500 new US installations of industrial robots in 2025. If tariffs land, those costs get paid by the US companies importing the robots, which means higher manufacturing costs at exactly the moment everybody's trying to reshore production. Now, there's an escape hatch, same one we've seen before, waivers. As the piece notes, Intel recently showed you can still get a waiver from the administration to keep using foreign robots even under a nominal ban. Which, if you've been listening to the top of this episode, should sound familiar. The formal rule, and then the discretionary exception for whoever's got the right relationship. Same pattern, different hardware.
Now let's move from the factory floor to the open web, because there's a story in the earnings pile that I think is genuinely a canary in the coal mine. Reddit reported Q3 earnings on Thursday, and buried in there is an announcement that should make every founder who depends on organic search traffic sit up. Reddit is shifting to a primarily "logged-in experience." And the reason is blunt: search traffic is drying up.
Here's the data, and it's stark. Reddit's US daily active users dropped from 53.2 million in Q2 to 50.7 million in Q3, down 2 percent year over year. Global daily actives were flat at 130.3 million. But the trend line is the scary part. The company said search traffic referrals dropped by more than a million referrals per week on average in October. And they warned that if the trend continues, daily actives could fall by a low-double-digit percentage next quarter. To put a number on it, a 10 to 12 percent drop takes them from 130 million down to roughly 115 to 117 million. That's a huge move for a company this size.
Why is this happening? Because people used to append the word "Reddit" to their Google searches to get real human answers. Now AI-powered search just summarizes the answer and you never click through. Google's own "ten blue links" matter less, AI summaries give you what you need without a visit, and this is hammering publishers and platforms across the board. It's even showing up in Google's own bottom line. So Reddit's move is defensive on two fronts. One, going logged-in protects its data from AI scraping, which lets Reddit instead sell lucrative licensing deals to the model providers. Two, logged-in users are worth more to advertisers because they throw off more signals.
And the revenue numbers back up the strategy, which is the wild part. Even with user growth weakening, revenue jumped 53 percent year over year to 895 million dollars. Net income rose 74 percent. US average revenue per daily user climbed 53 percent to 13.82. So they're getting way better at squeezing money out of a shrinking, more valuable audience. CEO Steve Huffman wrote, and this is the line that stuck with me, "I don't know about you, but as someone whose life has been defined by the open internet, it's a big change for me. But I've come to the conclusion that this is the right decision for our platform and the right decision for our business."
Here's why this matters for you as a builder, and it connects to something we've been circling all month. We've talked about the economics of the AI web, about Slok's "zero payoff" thesis, about how the model layer is reshaping value. This is the ground-level version of that. If you built a business on top of free organic discovery, Reddit just told you the model: that era is closing, and the platforms are walling their gardens so they can rent the data to the AI companies instead. Discord, by the way, is moving the opposite direction, making public servers more searchable starting November 16th, which sparked its own privacy backlash. So you've got platforms pulling in two directions, but the common thread is the same: your content, your data, is now a licensing asset, and the platforms are deciding who gets to see it and who has to pay. If your product depends on scraping, indexing, or free access to any of these walled communities, start planning for a world where that door is closed and there's a toll booth.
Now, let me bridge from the open web to the open models, because there's a genuine model release worth your attention, and it fits the theme of "who controls the stack." Mistral released the weights of Mistral Large 4, which they've nicknamed Le Chonk. It's a one-trillion-parameter multimodal model, and the pitch is squarely aimed at what they call sovereign AI, European-trained, fully open-weight, deployable on your own infrastructure. I'm going to keep this tight because the model-benchmark weeds are not where founders live, but the what and the why matter.
The what: Mistral says Large 4 is the best open-weight model for coding, leading on a couple of private benchmarks, and it does it with substantially better token efficiency, 47 percent fewer tokens per task than the comparable leading open model on one coding index. Translation for your budget: cheaper to run at scale. It's priced at 70 cents per million input tokens, 2.10 per million output. It's got three reasoning modes so you can dial cost versus performance. And they've leaned hard into defensive cybersecurity, with a long parade of security vendors, CrowdStrike, Cloudflare, Fortinet, Cisco, all saying nice things about running an open model they can audit and host themselves.
The why, and this is the strategic part: the whole sell is control. You can inspect it, customize it, run it without depending on Mistral's API. European banks like La Banque Postale and Crédit Agricole are in the quotes talking about technological sovereignty as a prerequisite for customer trust. And here's the one caveat for founders to actually read: the license is a modified MIT license, free to use, modify, and commercialize, with one condition. If you run a model-as-a-service offering, you need a separate commercial license from Mistral. So if your business is hosting and reselling the model, that's a conversation. If you're just building on top of it for your own product, you're clear. That's the fine print that matters.
And notice the through-line from the Reddit story. Everybody is fighting over who owns and controls the intelligence and the data. Reddit's locking its data up to license it. Mistral's giving the model away but gating the resale. The whole industry is drawing lines around what you own versus what you rent.
Now let's shift gears from models to the boring-but-critical stuff that keeps your shipped code from blowing up. Meta's engineering team put out a piece on something they call Catching JiTTests, and this one I actually want to spend a few minutes on because it's a real production result with real numbers, and it speaks to how AI changes the software development lifecycle in a way that's directly useful to builders.
Here's the setup. Meta has hundreds of millions of lines of code changing constantly. Traditional testing can't keep pace. So over the last year they've been building a system where large language models catch bugs on the fly, before code lands in production. And as of this piece from October 29th, it's not a research prototype anymore, it's part of Meta's infrastructure. Now, here's what makes their approach different, and it's a distinction worth understanding. Most AI bug-catching tools work like a faster human reviewer, the LLM reads your code change and flags what looks wrong based on prior experience. Meta's thing is different. It generates actual executable tests designed to fail, tests that demonstrate a concrete behavioral difference your change introduced. It's not opinion, it's proof. The test runs, it fails, and that failure is evidence of a real behavioral change.
They run two workflows in tandem. One they call "intent-aware," which tries to infer what your code change was supposed to do and targets tests at the likely failure points. The other they call the "dodgy diff" workflow, and I love the name because it tells you the philosophy. It treats every single code change as if it were a bug, a mutant, and generates tests to expose how it behaves differently from the parent code. It implicitly asserts your diff is buggy and tries to prove it. Guilty until the tests say innocent.
Then comes the filtering, because the challenge at Meta's scale is telling the real bugs, the "strong catches," from the false alarms, the "weak catches." They use an ensemble they call LLM-as-Judge, a collection of state-of-the-art models that look at the test code, the failure output, the actual diff, and the inferred intent, and vote on whether it's a true positive. They pair that with a rule-based deterministic system they call RubFake, because LLM judgments are non-deterministic and hard to verify, while rules give you a checkable answer. The two work together to rank candidates.
And here's the part I want you to really hear, because it's the organizational insight, not the technical one. Instead of "human-in-the-loop," where an engineer reviews every output, they went with "human-on-the-loop." The engineers set the criteria and review outcomes at an aggregate level, at the policy level, rather than babysitting every single result. Let me quote their framing of why it matters: "Combining these tests with a human-on-the-loop approach helps ensure that engineers rarely need to read test code and only need to confirm whether an observed behavioral change was intended." That's the whole game right there. The AI does the generation and the first-pass judgment, the human only gets pulled in to answer one question: did you mean to do that?
Now the results, because this is what separates a cool demo from a real thing. As of the writing, 589 engineers at Meta, which is 64 percent of the people they reached out to, have engaged with the system. Those engineers report a 60 percent true-positive rate on the findings, meaning six out of ten flags are real. And here's the headline number: in the first half of 2026, Catching JiTTests played a key role in catching 74 serious privacy bugs at the diff stage, before they hit production, and lowered the overall occurrence of privacy severe-incidents by 20 percent over that timeframe. Seventy-four privacy bugs that never reached a single user.
So why does this matter for you if you're not running Meta-scale infrastructure? Two reasons. One, the piece names the real problem of this era directly: agentic software development is letting everybody write and ship code faster than ever, which means you need to test faster than ever too. If you've got AI writing more of your code, you need AI watching that code more aggressively, because the humans can't keep up with the volume. The bottleneck moves from writing to verifying. Two, the human-on-the-loop pattern is the generalizable lesson. The quote worth stealing: "Engineers define success criteria early, let AI execute the testing, monitor the outcomes at an aggregate level, and intervene at the policy level rather than per item." That's a template you can apply to a dozen workflows in your own company, not just testing. Set the policy, let the agents run, review in aggregate, intervene by exception. That's how you actually get leverage out of AI instead of just adding another dashboard nobody reads.
And it connects to that spy story from the top in a funny way, actually. Meta's whole design assumes the agent might do something its developer didn't anticipate, so they enforce controls on what the agent attempts at runtime, not just what you expected it to do. The access decision doesn't depend on the agent behaving. That's the right mental model for any autonomous system, whether it's a testing agent or a geopolitical risk: don't trust that everyone plays by the script, build the guardrails to catch the off-script behavior.
Alright, let me run through a few more that deserve a mention before we close out.
On the earnings front, Amazon reported Thursday and it's a tale of two numbers. The bright spot: AWS, the cloud unit, grew a whopping 39 percent year over year, beating expectations, with sales hitting nearly 46 billion dollars and operating income jumping to 17.8 billion from 11.4 billion a year ago. That's the AI demand story in black and white. The ouch: free cash flow fell to negative 42.4 billion dollars for the trailing twelve months, compared to positive 14.8 billion a year earlier. They're burning cash, driven by AI investment. Amazon still expects to end 2026 with about 220 billion in capital expenditures, largely for AI. To put that in perspective, that's almost as much as they expect to bring in from sales in the entire fourth quarter, the biggest quarter of the year thanks to holiday shopping. So the cloud business is booming and the capex is staggering, all at once. That tension, roaring AI revenue against brutal cash burn, is the whole hyperscaler story in one earnings report.
Quick one on OpenAI, which updated its Model Spec, the document that outlines the core values guiding how its models behave, now built around five core values aimed at sound judgment, human agency, and faithfulness to human intent. They also put out a piece on how they detect frontier model misalignment, layered monitoring to catch unauthorized agent behavior before incidents escalate. Worth a read if you're building agents yourself, because the monitoring philosophy, catch the bad behavior before it escalates, is the same instinct we just saw in Meta's testing work.
And on the EV front, a couple of quick data points for anybody tracking that market. Rivian said its new R2 SUV, which launched back in June, is already its best-selling vehicle and pushed the company to raise delivery guidance for the second time this year. The notable bit: Rivian is finding buyers despite the death of the 7,500-dollar federal EV tax credit, in a market where a lot of automakers are retreating back to gas and hybrids, leaving room for the pure-EV players. And in a related note, Rivian's expanding its joint venture with Volkswagen to include licensing its in-house AI voice assistant, the one built on its own agentic framework that could eventually run agents from other companies. The vertical-integration bet is paying off, VW's buying the software because it couldn't build a modern software-centric car on its own.
And one last human-interest nugget before we go, because this episode's been heavy on policy and geopolitics. Airbnb's nonprofit threw its first-ever gala in Los Angeles on Wednesday, hosted by Jimmy Kimmel, Charli XCX performing, Serena Williams and Charlize Theron on the guest list, and the entire internet's reaction was the same three words: "a what gala?" The actual news underneath the confusion is that Airbnb's donating ten cents of commercial revenue per stay to its disaster-housing nonprofit, which it says adds up to 500 million dollars over five years. It's a tidy reminder that a splashy party is still one of the best ways to get rich people in a room and get media attention, even if nobody can quite believe the invite list.
So that's the menu today, kiddos. A spy plot that may be steering chip policy, a robot tariff aimed at machines we can't build ourselves, Reddit walling off the open web to sell its data to the AI companies, a European open model with a resale catch, and Meta proving that the way to survive AI-speed coding is AI-speed testing with humans watching from the policy level. The thread, if you want one: everybody's figuring out who controls the stack, the data, and the rules, and the rules on paper matter less than who's in the room. Build accordingly.
This is Tony DeLuca. Watch the stuff that outlives the headline, be good to each other, and I'll catch you on the next one.