Super Intelligence News and Report — Ground Stop: Did Aging Air Traffic Networks or the New AI Tool Cause Nationwide Shutdown?

Welcome to Super Intelligence News and Report. Today is Saturday, November 14th, here are the headlines.

Maya: Today's thread is old plumbing under strain: upgrades bolted onto aging systems, and a chip supply chain still shaking underneath AI. First, Nvidia closed out Jensen Huang's last week down about thirteen percent. It slipped another two percent on Friday, the day after Huang resigned as CEO, and finished below a four trillion dollar market value for the first time since July of last year. HSBC cut the stock to hold until there's a permanent succession plan, and Reuters reported another affair, this one with an unnamed Nvidia employee.

Theo: How long has the slide run?

Maya: Six straight weekly losses, the longest streak since 2012, while AMD and Broadcom both rallied hard. Interim CEO Colette Kress, who's still CFO, reports earnings Tuesday, November 17th. The market's bet is that customers start shopping elsewhere. And the scandal is spreading. Bloomberg reported that a Shanghai firm founded by Alice Huang, the former Megaspeed executive at the center of the Nvidia scandal, ended up with servers holding H800 chips, which the US banned for export in 2023. That's reported, not confirmed. The same day, China's foreign ministry warned Washington against new export curbs, and TechCrunch says Commerce Secretary Howard Lutnick could lose his job over his department's handling of Megaspeed. This has moved from boardroom conduct to whether export controls actually work.

Theo: Which leads straight into some clever detective work. Will DePue, a former OpenAI researcher who now runs a nonprofit called Unemployment Labs, went through the released weights of open models from DeepSeek, Moonshot, MiniMax and Ant Group. He says he found compelling evidence that Kimi K3 and DeepSeek V4 were trained at least partly on Nvidia Blackwell GPUs. It's reported, not confirmed: one researcher, via TechCrunch, and he hasn't published his code.

Maya: Does that mean somebody broke the law?

Theo: Not necessarily. The chips could sit in offshore data centers, Malaysia for example. But Epoch AI estimates about two million advanced chips reached China across 2024 and 2025 through routes that hid their destination. What's new is the method: auditing where a model was trained by reading its weights. On the product side, OpenAI launched ChatGPT Wallet on Friday for a limited group of US Pro subscribers, eighteen and over. You add a Visa, Mastercard or Amex once, OpenAI's agents, which it calls dots, go shopping, and you approve each purchase with Face ID or a fingerprint. A payment token stands in for the card, so neither the agent nor the merchant sees your number. No retailers or fees disclosed yet.

Maya: A checkout button with nothing named to check out from.

Theo: For now. But it's a clear push on agent checkout while Amazon and others are blocking agents outright.

Maya: Back to plumbing. Microsoft says a bug introduced during a network capacity upgrade caused its Outlook, Teams and Copilot outage. It started around one p.m. Eastern on Thursday, ran more than a day, and eased Friday. Servers dropped offline and their load swamped the rest, with Downdetector peaking around two hundred thousand reports. Microsoft says it wasn't a cyberattack or AI agents, apologized, and promised a full root-cause report. Copilot's enterprise pitch rests on uptime, and Ignite starts Tuesday. On the legal front, Ontario's Court of Appeal dismissed OpenAI's bid to move the Canadian publishers' copyright case to a US court. The plaintiffs include The Canadian Press, The Globe and Mail and CBC/Radio-Canada, who say OpenAI trained ChatGPT on their journalism without permission or payment. It's reported as the first Canadian case on training AI with copyrighted work, and alongside a similar suit against Cohere, it gives AI copyright fights a second national arena.

Theo: Now a promise finally kept. Thinking Machines says the Inkling 2 Small weights are on Hugging Face under Apache 2.0, after the closed launch on November 5th and a 'next week' that slipped. It's a mixture of experts with seventeen billion active parameters out of nearly six hundred billion total. That's from the company's announcement, so reported, not confirmed independently.

Maya: Can anyone actually run it?

Theo: Not on a laptop. The main checkpoint needs a GPU cluster with at least five hundred gigabytes of combined memory. Still, a US lab now has a big open-weight model on the board. Next, a breach. Notion says a breach at its inference provider, Fireworks AI, exposed AI chat transcripts from its Insights feature in under one percent of workspaces, and most couldn't be tied to a user. Notion cut off Fireworks and turned Insights off. Fireworks' own notices describe a second intrusion that exposed billing data and, for some accounts, secrets, datasets and incomplete model weights.

Maya: So your vendor's vendor is your problem now.

Theo: It's part of your attack surface, yes. And another agent stepping into the real world: Meta's chief AI officer Alexandr Wang says paid Muse subscribers in the US can now have the agent call insurers, cable companies, the DMV and restaurants. He says Muse has more than ten million users. The catch is that plenty of businesses hang up once they realize it's an AI. Put that next to Wallet: agents are acting in the world, and the world is pushing back.

Maya: Some money for Nvidia's would-be rivals. Fractile, based in London and Bristol, closed a seven hundred twenty million dollar Series C, led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, at a reported valuation of about seven billion. Its inference chip skips off-chip HBM memory, and the first commercial parts aren't due until 2027.

Theo: Seven billion, and the chip hasn't shipped.

Maya: That's the point. Investors want Nvidia alternatives badly enough to pay up years ahead of revenue.

Theo: And a hidden gem from the changelogs, tracked by third parties, so treat it as reported. Claude Code added an experimental, opt-in Code Mode, where Claude writes JavaScript to call its tools instead of calling them one at a time. Anthropic also opened a private beta for running Managed Agents with session data kept on a company's own servers. And OpenAI's Codex CLI added model-provider switching and multi-tab side chats. The coding tools are converging on agents that write code to drive their own tools.

Maya: Next, we go deep on Friday's nationwide ground stop, three days after the FAA took its new AI traffic tool to every center: was it the AI, the old network, or the mix?

Now for the Main Episode

Maya: Here's the question. On Friday afternoon, every departing flight in the country was held on the ground, three days after the FAA switched on a new AI traffic tool at all of its en route centers. So was it the AI, the decades-old network underneath it, or something new bolted onto that network that the old wiring couldn't carry? And whichever it was, what does it tell us about putting AI into safety-critical systems that still run on copper?

Theo: Start with the tool, because I think a lot of people are picturing something it isn't.

Maya: Go ahead, that part's yours.

Theo: It's called SMART, short for Strategic Management of Airspace, Routes and Trajectories. It takes in weather, flight paths, traffic and controller staffing, forecasts where congestion is building, and suggests more efficient routes. The key point is that it doesn't control aircraft. It isn't talking to planes or replacing anyone on a radar scope. It advises on traffic flow, the big-picture layer where someone decides to meter arrivals into Atlanta because a storm line is coming.

Maya: And the money behind it. The vendor is Air Space Intelligence, ASI, which won an eight hundred seventy-five million dollar contract in June, running twelve years. That's roughly seventy-three million a year, after twelve companies bid. Meaningful for the FAA, but not a moonshot.

Theo: One detail from the trade press matters. SMART is an add-on inside a larger ASI platform that's meant to become the technology backbone of the FAA's national Command Center. So the AI piece is the visible tip of a bigger rebuild.

Maya: Now the timeline, which is the whole reason we're asking. SMART went live around Washington, D.C. in September. Then it went to all twenty-two air route traffic control centers. KCRA and others put that on November 10th. Gizmodo, as cited by Jalopnik, says closer to November 12th. Either way, days before the outage, not weeks. And Transportation Secretary Sean Duffy announced the twenty-two-center activation himself.

Theo: Which makes his later denial land a little differently.

Maya: We'll get there. Friday afternoon, communications problems hit air traffic control centers across the country. Controllers reported frequency issues in multiple centers. In one recording out of Atlanta, a controller says he's been told to ground stop everything. The FAA ordered a nationwide ground stop on all departing flights, citing a telecommunication outage and extraordinary circumstances.

Theo: How big did it get?

Maya: Flightradar24 counted about thirty-one hundred aircraft already in the air and headed for US airports when the stop began. Those planes keep flying, so you've got a sky full of inbound traffic and a network struggling to talk to itself. Duffy says the airspace reopened around quarter past eight Eastern, after about five hours.

Theo: And the comparison is the NOTAM failure.

Maya: Right. January 2023, when the pilot-notice system went down, and that ground stop lasted under two hours. This one ran more than twice as long. Before 2023, the last nationwide stop was after 9/11. Robert Sumwalt, the former NTSB chair who now analyzes this for CBS, called it absolutely unprecedented.

Theo: And the knock-on?

Maya: FlightAware listed more than ten thousand delays Friday and more than four thousand cancellations, with some reports saying forty-seven hundred. CNN cited FlightAware expecting about three thousand more flights hit today, because planes and crews ended the night in the wrong cities. Five hours on Friday becomes a weekend problem.

Theo: So, the explanations. My read is nobody official is telling the same story.

Maya: They aren't, and that's the most interesting part. Duffy went first and was categorical. He blamed old equipment: an outdated analog system, the TDM kind, plus switches, routers and copper lines that got overloaded with data and started producing inconsistent information, so they shut it down. On SMART, he said it isn't plugged in to the core radar and flight-tracking systems, and there was, quote, no possible way it caused this.

Theo: That's a very strong claim to make the same evening.

Maya: It is. Then Homeland Security Secretary Markwayne Mullin. No sign of a malicious attack, he said, but he didn't rule one out. On AI, he said we can't rule out anything at this point, and investigators were looking at the rollout. Then the line I think matters most: as the agency added patches, new equipment and new technology to the system, those appeared to overwhelm the older routers.

Theo: New equipment and new technology. Did he say what?

Maya: No. Neither did Duffy.

Theo: That's the gap I keep coming back to. Duffy says the old routers choked on data. Mullin says they choked on new stuff. Those could both be true. But 'the AI couldn't possibly be involved' and 'new technology overwhelmed the routers' don't sit together easily until someone names the new technology.

Maya: Then there's a third account, the most specific one. CBS News, citing multiple unnamed sources, reported that a software update to the private network linking the air traffic control system triggered a reboot, then a crash, then the communications outage. That network is run by the defense contractor L3Harris, which hadn't responded to requests for comment. That's reported, not confirmed. On the same thread, SecurityWeek reported that the FBI is examining whether a software update was responsible, and has found no evidence of hacking.

Theo: As an engineer, that's the account that sounds right to me. An update goes out to a shared network, something reboots badly, the old gear around it can't absorb the disruption, and it cascades. Very ordinary failure. What's ASI saying?

Maya: ASI posted on X that the ground stop was not based on or a result of SMART, any ASI product, or its integration with the FAA. And Mullin said there was no link to the Microsoft outage.

Theo: Good, because Microsoft is the useful comparison here. The day before, a bug introduced during a network capacity upgrade knocked servers offline, their load piled onto the rest, and Outlook, Teams and Copilot were down for more than a day. Microsoft explicitly said cyberattacks and AI agents weren't the cause. Two big cascades in two days, and in both the trigger was a change to the network, not the AI running on top of it.

Maya: I buy the pattern. But Microsoft's explanation came from the company that owns the whole stack. At the FAA, the people explaining it don't fully agree, and one of them announced the tool's rollout three days earlier.

Theo: That's fair.

Maya: And that's the skeptics' case. Daniel Golson at Jalopnik openly mocked how confident the official denials were. To be fair to him, he says plainly that nothing concrete ties SMART to the failure. His point is that the timing is hard to ignore, and the Homeland Security secretary himself wouldn't clear it. Futurism framed it the same way, as an open question of coincidence.

Theo: Here's where I'd push back a little. 'Not plugged in to core radar and flight tracking' is a specific, checkable engineering claim. An advisory flow tool should sit off to the side of the safety-critical loop. That's how you'd design it. If that's true, SMART can't reach in and break the frequencies controllers were losing.

Maya: But does it have to reach in to cause trouble?

Theo: No, honestly. A tool like this lives on data. Weather, flight paths, staffing, all moving across the network in real time. Not plugged in to radar isn't the same as putting no traffic on the routers. Light up a data-hungry system at twenty-two centers in one week, over routers a 2023 review already called obsolete, and you've added load to the exact part Duffy says gave out. Nobody has said whether SMART's feeds touched the affected systems.

Maya: So Duffy's own explanation makes it harder for him to rule SMART out completely.

Theo: A bit, yes. 'The old routers were overloaded with data' and 'the new data-heavy tool couldn't possibly matter' is an awkward pairing. But I still think the most likely story is the network update CBS is pointing to. It's specific, the FBI is reportedly looking at exactly that, and it fits the Microsoft pattern.

Maya: I'm less settled than you, and not for technical reasons. Look at what Duffy did with the outage. Almost immediately he asked Congress for about fifty billion dollars in emergency funding for a full air traffic control replacement.

Theo: Against what's already been spent?

Maya: Coverage cites five billion from the 2021 infrastructure law and twelve and a half billion from the One Big Beautiful Bill, so about seventeen and a half billion. The new ask is nearly three times that. And it's a request. Congress hasn't appropriated anything.

Theo: Aren't those consistent, though? 'The old system is obsolete, fund a replacement' and 'the new tool is fine' both point at the old network.

Maya: They are, and that's my point. Blaming the old wiring clears the tool he launched and supports the check he wants. It may also be true. CBS noted an FAA-commissioned safety review back in November 2023 had already called the network obsolete. And Jalopnik reported, days before this, that the FAA was still phasing floppy disks out of air traffic control.

Theo: Floppy disks, the same month it rolled out an AI routing platform.

Maya: That's the image for the whole story. Duffy isn't wrong that it's old. I just think a secretary with a preferred outcome shouldn't be the one declaring certainty on day one. And carefully, there's one more item. Raw Story reported that ASI gave twelve thousand five hundred dollars to the congressional campaign of a Duffy relative, and framed it as a conflict of interest. Reported, not confirmed, one outlet's reporting and framing. I wouldn't build anything on it. But it's another reason an independent root cause matters more than a cabinet secretary's word.

Theo: Is Congress pushing?

Maya: So far in statements. Senator Tammy Duckworth, the top Democrat on the Senate aviation subcommittee, says she'll demand answers from the FAA. Senator Richard Blumenthal wants a prompt, thorough investigation. Representative Robert Garcia wants transparency from the FAA, the Transportation Department and the White House. Duckworth's the one to watch, because a subcommittee hearing is where someone could be asked what Mullin's new equipment actually was.

Theo: For anyone selling AI into government or critical infrastructure, ASI just learned that being the newest thing in the building makes you the first suspect. That's not unfair. If you're the most recent change, you get investigated. And going from one region in September to all twenty-two centers within a few days is fast for anything touching national airspace, even advisory. A slower rollout gives you a cleaner before-and-after. The real risk usually isn't the model making a bad call. It's the integration: the data pipes and the load on infrastructure you don't own and can't see.

Maya: And for ASI as a business, a twelve-year contract is being tested very publicly in its first months. If the root cause clears it, it probably comes out stronger, the modern piece that wasn't the problem. If SMART's traffic was even part of the load, the politics change a lot, because Duffy staked his credibility on no possible way.

Theo: So where do you come out?

Maya: Three things. First, from what's reported, the likeliest cause is ordinary: a change to an old network that couldn't handle it, whether the L3Harris update, Mullin's new equipment, or both. Second, 'it wasn't the AI' and 'new technology overloaded the system' are both coming from the same government, and they can't both be the full picture until someone names the equipment, and until L3Harris gives its side. Third, the fifty billion is the story to follow. It may be the right fix, but it's being argued from an outage whose cause hasn't been established.

Theo: I'm close, but I land a bit differently. I think the AI question is mostly a distraction from the bigger risk: the FAA is adding modern software to infrastructure that can't safely absorb change of any kind. AI, a patch, a router swap. Microsoft showed the same thing with no intelligence involved. Where I agree with the skeptics is that the FAA owes everyone a yes or no on whether SMART's data load touched those routers. That should be answerable.

Maya: And the clearest signal of whether they believe their own denial?

Theo: Whether SMART stays running at all twenty-two centers. If the FAA quietly pulls it back to Washington while the investigation runs, that tells you more than any statement. If it keeps running everywhere and the root-cause report names the L3Harris update, Duffy's call holds up.

Maya: I'll be watching for a published root cause from the FAA, the Transportation Department or the FBI, whether Duckworth gets her hearing, and whether the fifty billion moves or just sits there as a talking point.

Theo: So, was it the AI? Probably not, from everything we can see. But that's not much comfort. The more worrying answer is that it doesn't take AI to ground the country. A routine change to a network this old can do it, and every time you add something smart to that system, you're betting the old pipes can carry it.

Maya: And on Friday, at least one bet on the old pipes failed. We just don't know yet which one.

Thanks for listening to Super Intelligence News and Report. Please subscribe and share it with your friends.