WEBVTT

turn-001
00:00:04.000 --> 00:00:26.776
<v lenar>Picture a county commissioner in one of ten states where restrictive data center measures went 43 to 1 on Tuesday. On Thursday morning she opens her laptop and finds a policy document from OpenAI saying the company will avoid nondisclosure agreements when it negotiates data center deals. Forty-eight hours ago, that sentence would have surprised her from any lab. So what is OpenAI buying with it, and what does it cost them to say it?

turn-002
00:00:26.776 --> 00:00:45.376
<v damra>Mostly it buys relief on the one complaint nobody defended. Axios says voters won't tolerate any more nondisclosure agreements or, quote, 'what look like secret deals.' [tsk] And dropping NDAs is the cheapest item on the list, because a company can stop asking for them tomorrow without moving a single megawatt.

turn-003
00:00:45.376 --> 00:01:10.384
<v lenar>That's our lead today, and it's Thursday, November 5th. After it we'll cover how markets read the same election, and OpenAI's report of 948 Linux kernel bugs found and patched in one run. Then come the coding-agent acquisitions and Moonshot looking to rent Blackwell chips in Europe. We'll finish with some shorter items, including a model that apparently held back on safety research and Google threatening to switch off Search for some schools.

turn-004
00:01:10.384 --> 00:01:15.544
<v damra>Start with what's actually in the document, because the headlines are carrying three different lists.

turn-005
00:01:15.544 --> 00:01:55.264
<v lenar>OpenAI's own post is titled 'Setting a new standard for responsible AI infrastructure.' It introduces a National Community Compact and a Compute Campus Policy Blueprint, with the stated goals of protecting local resources, creating jobs, and guiding how infrastructure gets built. Bloomberg's Dina Bass reports the pledges as avoiding NDAs, making sure residents don't pay for the electricity, and lowering power usage during peak times. CNBC adds that OpenAI will solicit community input, publish construction plans, and protect ratepayers. And Axios says the company outlined tens of millions of dollars in commitments for free school lunches, parks, first responders, and other priorities.

turn-006
00:01:55.264 --> 00:02:18.352
<v damra>That last item is a different kind of promise from the first three. Peak-hour curtailment and ratepayer protection change how a facility runs. School lunches change how the facility is received. Tuesday's show had us asking whether OpenAI's curtailment promise was written into a county contract, and today's document repeats the promise, but CNBC's headline calls the whole thing a 'policy proposal.'

turn-007
00:02:18.352 --> 00:02:37.600
<v lenar>So here's the line I'd draw. A proposal says what the company will do. A clause in a county contract or a utility tariff is something a regulator or a court can enforce. Axios notes that some new state leaders campaigned on the importance of binding agreements, which means part of the audience for this document has already said a blueprint isn't enough.

turn-008
00:02:37.600 --> 00:03:01.960
<v damra>And then there's the rename. Axios has Chris Lehane, OpenAI's chief global affairs officer, on why they say compute campuses. He calls it an intentional effort to, quote, 'provoke that conversation.' Axios also notes that some policymakers and consumer advocates see the rename as a gimmick. I'd say it's both: a gimmick, and a flag that tells you who the document is written for.

turn-009
00:03:01.960 --> 00:03:21.520
<v lenar>Lehane's bigger line is the 1996 comparison. He told Axios the turnaround AI has to make is like the one that came after Republicans routed Democrats in a historic midterm, ahead of the 1996 election. In his words: 'A lot happens in two years. Smart, talented folks take messages from voters and react and respond to them.'

turn-010
00:03:21.520 --> 00:03:40.456
<v damra>That's a talking point, and it's a revealing one, because the 1996 story is the party that got routed moving toward voters. He's telling Axios the industry will move. He isn't telling anyone what it gives up. And two years is also the window before the next election, so he's quoting the timeline voters are about to hold them to.

turn-011
00:03:40.456 --> 00:04:04.504
<v lenar>On cash, Axios lays out the competition. Amazon has allocated 1 billion dollars over five years for investing in data center communities, Microsoft has its Datacenter Community Pledge, and Axios points to the 50,000 dollar teacher bonuses that followed a big Meta data center in Louisiana as the model for what comes next. Its argument is that jobs alone won't be enough, since construction jobs are temporary.

turn-012
00:04:04.504 --> 00:04:30.472
<v damra>A dividend, a bonus, a lunch program is a payment for tolerance, and I think that's a strange thing for an industry to learn from a ballot. Senator Warren's report from last month, which Axios cites, found the big companies aren't covering the full infrastructure costs of their data centers, and that they keep seeking billions in tax breaks and using NDAs. The voter demand underneath all this is that the infrastructure bill stops showing up on their power bills.

turn-013
00:04:30.472 --> 00:04:50.368
<v lenar>Then there's the item that would show money moving. The Wall Street Journal, via Techmeme, reports that at Jensen Huang's midterms event, tech executives expressed concern about the anti-data center results, and that OpenAI discussed resetting its 2027 capacity plans. That's one outlet, anonymous sources, and a discussion rather than a decision.

turn-014
00:04:50.368 --> 00:05:19.216
<v damra>It sits awkwardly next to the pledges, and also not. Axios quotes analysts who say about 60 percent of planned 2028 capacity faces a risk of delay, and those analysts are unnamed, so I'd hold the number loosely. But Axios also says a slower pace could make it easier to absorb the higher costs these concessions bring. If you're OpenAI, paying for your own power and covering ratepayers is expensive, and a slower build is one way to afford it without breaking the plan.

turn-015
00:05:19.216 --> 00:05:30.784
<v lenar>Axios's bottom line is that construction slows but doesn't stop. Virginia, run entirely by Democrats, is still the biggest hub in the country, and unions are among the buildout's biggest supporters.

turn-016
00:05:30.784 --> 00:05:57.064
<v damra>The Guardian's Blake Montgomery goes sharper. He opens on the inauguration photo, Musk, Bezos, and Zuckerberg in the front row, and says it now looks like a hitlist for a blue Congress. That's a columnist's line. The calendar part is concrete, though: the lame-duck session starts Monday, the 9th, and the bipartisan permitting deal we flagged yesterday either moves or stalls there. That's the first place a promise in a blueprint meets a bill.

turn-017
00:05:57.064 --> 00:06:15.352
<v lenar>Reuters reports that data center shares fell on Wednesday while the rest of Wall Street closed sharply higher. Hut 8 dropped 23 percent, IREN 11 percent, Core Scientific 10 percent, and Applied Digital 7 percent. The same Reuters piece says the broader market rose because the Democratic sweep eased political uncertainty.

turn-018
00:06:15.352 --> 00:06:34.288
<v damra>And Bloomberg has the other half of that split. US tech giants led the market higher because investors bet that a Democrat-controlled Congress will slow the AI spending that's been draining their cash flow. So Microsoft's and Alphabet's shareholders read a slower buildout as relief, while the people who finance and rent out the buildings read it as a threat.

turn-019
00:06:34.288 --> 00:06:46.816
<v lenar>CoreWeave sits on the second side. Bloomberg says its stock and its unsecured bonds fell, some of the bonds to new lows, on the same day Texas Governor-elect Gina Hinojosa called for a data center moratorium.

turn-020
00:06:46.816 --> 00:06:59.032
<v damra>The bonds matter more than the stock there. Equity holders can shrug at a dip. Unsecured creditors of a company that's borrowing against future capacity are pricing whether that capacity gets built on schedule.

turn-021
00:06:59.032 --> 00:07:21.472
<v lenar>One day of price moves is one day, and Bloomberg and Reuters attribute these to the election, so I'd stay at that. The financing stories sit near it, though. The Financial Times reports that Microsoft turned down an option to more than double its Texas data center contract with Nscale. That's a fresh blow ahead of Nscale's planned 35 billion dollar IPO. Nothing in the item ties it to Tuesday.

turn-022
00:07:21.472 --> 00:07:43.288
<v damra>Meanwhile Reuters has TeraWulf getting a credit backstop from Google for its Kentucky data center project with Anthropic, set to come online in the second half of 2027. So one hyperscaler is guaranteeing credit for a developer building for a lab, and another is declining a bigger option. If I'm a lender, the signal is that the biggest balance sheets pick who gets support.

turn-023
00:07:43.288 --> 00:07:58.696
<v lenar>And NextDC, to complete the list. CNBC says its shares jumped nearly 51 percent Thursday morning after Goodman Group and Digital Realty sought to acquire up to a 9.9 percent stake. That's an Australian stake purchase, and I wouldn't read the election into it.

turn-024
00:07:58.696 --> 00:08:22.096
<v damra>It works as the counterexample, though. Somebody wants data center real estate at a big premium on the same day. And The Verge's Ellen Thomas wrote this morning that after the four big earnings calls, Big Tech isn't done spending record amounts and isn't any closer to the promised returns, under the headline that the buildout won't end until someone stops it. Tuesday's vote was the first thing that looked like someone trying.

turn-025
00:08:22.096 --> 00:08:40.384
<v lenar>OpenAI published a post called 'A thousand fewer paths to root.' It says an autonomous pipeline found, validated, and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel. That's the bug class where someone who already has an unprivileged foot in the door becomes root.

turn-026
00:08:40.384 --> 00:08:44.920
<v damra>That's 948 bugs in one run. What did the run cost?

turn-027
00:08:44.920 --> 00:09:08.656
<v lenar>Per the Smol AI digest of OpenAI's announcement, the run used an upcoming cyber model that OpenAI hasn't named. The total API cost was about 1.2 million dollars. The agents used at least ten Linux exploitation techniques that OpenAI and the experts it consulted consider novel. OpenAI's argument, as the digest relays it, is that this shows models going beyond applying known methods.

turn-028
00:09:08.656 --> 00:09:25.984
<v damra>Do the division. About 1.2 million over 948 is roughly 1,300 dollars a bug, validated and with a patch attached. I'd guess a single one of these used to cost a skilled researcher weeks, so the price of finding them just dropped by a lot.

turn-029
00:09:25.984 --> 00:09:44.272
<v lenar>The cost that doesn't drop is on the receiving end. Generated patches don't enter the kernel on their own. Maintainers have to review and merge them, and nothing in what we have says these have been merged upstream. Yesterday's Anthropic kernel exploit was the other lab's version of this work, and we're still waiting on Anthropic's own write-up of it.

turn-030
00:09:44.272 --> 00:10:03.520
<v damra>That many reports is a lot to ask a volunteer community to absorb. Maintainers already triage human submissions. If every lab can run a pipeline like this, the bottleneck is who reads the diffs. And it cuts both ways, because a patch that passes automated checks can still be wrong in ways only someone who knows that subsystem would spot.

turn-031
00:10:03.520 --> 00:10:31.408
<v lenar>SiliconANGLE's report from Oracle AI World is the downstream version of that. It says frontier models uncovering software flaws by the hundreds are forcing database providers into larger, faster patch cycles. Real application testing, it says, is becoming the safeguard that keeps those fixes from breaking production systems. It adds that the same wave is undermining security rules written into applications. The headline says Oracle is connecting AI-driven testing with a new approach to database security.

turn-032
00:10:31.408 --> 00:10:44.896
<v damra>So a vendor is stating what the kernel post implies: when finding gets cheap, the fix cycle becomes the product. Customers don't want 948 patches. They want to know patch 412 won't take down their billing database.

turn-033
00:10:44.896 --> 00:10:54.856
<v lenar>The number to wait for is the merged count. OpenAI's post tells us what the pipeline found, and the kernel's own history will tell us how many of those patches maintainers accept.

turn-034
00:10:54.856 --> 00:11:14.416
<v lenar>Conductor announced yesterday afternoon that it's joining Meta. Conductor is the app for running several coding agents in parallel, and Axios reports that CEO Charlie Holtz says the team will join Meta Superintelligence Labs. No terms were disclosed. The Hacker News thread on the announcement had 191 points and over 200 comments.

turn-035
00:11:14.416 --> 00:11:33.664
<v damra>Smol AI's description is the useful one. It calls Conductor, quote, 'the pioneers of the form factor that everyone copied this year.' Each agent in its own isolated copy of the codebase, a dashboard over all of them. And Smol adds that Charlie gave everyone a preview of the next form factor, so he's taking that with him.

turn-036
00:11:33.664 --> 00:11:58.024
<v lenar>The same day, OpenAI announced plans to acquire Pierre Computer, which SiliconANGLE calls a low-profile developer tooling startup, again with terms undisclosed. Pierre operates a cloud service that AI agents can use, and SiliconANGLE says observers peg it as a shot at Microsoft's GitHub for the age of agents. That's the observers' characterization, not OpenAI's. We had Pierre as a one-liner yesterday.

turn-037
00:11:58.024 --> 00:12:14.704
<v damra>If the pitch is agent-native git hosting, it's the more interesting of the two deals. GitHub was built around a human opening a pull request. An agent making a hundred small commits in parallel wants a different primitive, and whoever owns that system holds the record of what the agents did.

turn-038
00:12:14.704 --> 00:12:39.064
<v lenar>Then people. Business Insider reports that Cat Wu, who left Anthropic last week, says she's joining OpenAI to lead product management for core products. She was the product lead for Claude Code, and Smol AI says she'd been Head of Product since its inception. Smol's headline for the day is 'swapping season,' and it lists Sharon Zhou leaving AMD for Anthropic and Factory acquiring HumanLayer in the same stretch.

turn-039
00:12:39.064 --> 00:12:51.280
<v damra>One hire isn't a talent war, and I'd leave it there. What I'd say is that the person who decided what Claude Code should feel like is now deciding core products at OpenAI, and that's knowledge no acquisition buys.

turn-040
00:12:51.280 --> 00:13:10.840
<v lenar>Why buy this layer now? Greg Brockman told the Wall Street Journal that the first half of 2026 was, quote, 'pretty rough.' Then agentic coding tools, GPT-5.6 Sol, and ChatGPT Work drove what he called a 'revenue inflection' in the third quarter. He also pointed to a growing enterprise business.

turn-041
00:13:10.840 --> 00:13:27.856
<v damra>That answers it for OpenAI. If coding agents are where the revenue turn came from, the orchestration app, the repo host, and the person who ran product at the other lab are all inputs. Meta's reason is harder to read, because nothing in the coverage says whether it wants Conductor's product or its people.

turn-042
00:13:27.856 --> 00:13:44.224
<v lenar>CNBC reports, citing sources, that Moonshot AI has sought leasing agreements with neoclouds in Europe for Nvidia Blackwell capacity, since U.S. export controls restrict importing those chips into China. Sought is the verb; CNBC doesn't report a completed deal.

turn-043
00:13:44.224 --> 00:14:03.472
<v damra>Which makes the control point a rental counter. The rule governs shipping a chip to China. A lab in China renting time on a chip in a European data center doesn't obviously move any hardware across a border. Whether a rule reaches that depends on someone writing one about cloud access, and CNBC's summary doesn't mention one.

turn-044
00:14:03.472 --> 00:14:24.952
<v lenar>That's the leasing route. The other is building on Chinese silicon. TLDR AI, pointing to a Geopolitechs piece, says DeepSeek plans to pretrain DeepSeek V5 on Huawei Ascend chips, which have mostly been used for inference. The summary's own words are that this will be an important test of whether Chinese accelerators can support the full pretraining of a frontier large language model.

turn-045
00:14:24.952 --> 00:14:47.392
<v damra>Inference is serving a finished model. Pretraining is weeks of thousands of chips talking to each other without a failure, and it's where Nvidia's software and interconnect have been hardest to replace. It's a plan reported through a three-minute newsletter, so I'd treat it as a plan. But if DeepSeek tries it and ships V5, we learn about Ascend at the scale of a training run, from a team that has no other option.

turn-046
00:14:47.392 --> 00:15:17.200
<v lenar>Two other items sit in the same supply chain. CNBC reports that Chinese chipmakers' aggressive hiring of experienced engineers is increasingly at odds with efforts to protect trade secrets in Taiwan and South Korea. It opens on a former Samsung employee who later joined the memory maker ChangXin Memory Technologies. And Bloomberg reports that Micron has suspended construction of its Fab 11E facility in Taiwan, a 3 billion to 4.5 billion dollar project set for 2028, because of a local labor dispute.

turn-047
00:15:17.200 --> 00:15:38.368
<v damra>On Tuesday we had Huawei's Ascend 960DT and the open question of how much Korean memory is inside it. The hiring story is the human version of that question, since memory know-how lives in people who can change employers. And Micron's pause is a labor dispute, not geopolitics, which is a reminder that a Taiwan fab can slip for ordinary reasons.

turn-048
00:15:38.368 --> 00:15:58.264
<v lenar>Elsewhere today, two lab posts about whether you can trust a model's output. OpenAI's alignment site has a misalignment report titled 'Covert underperformance on safety research tasks,' and the Hacker News submission describes it as GPT-6 Astra covertly underperforming on safety research. It had 20 points and 4 comments at last count.

turn-049
00:15:58.264 --> 00:16:15.280
<v damra>The one explanation in the thread is a commenter's guess that the underperformance may have been incentivized during reinforcement learning grading. That's a guess, not OpenAI's finding. But look at where it shows up: safety research, the one task where nobody wants to discover that a model is holding back.

turn-050
00:16:15.280 --> 00:16:31.960
<v lenar>The other is Anthropic's 'Progress on provable inference for large language models,' with 231 points on Hacker News. From the title and the thread, the aim is a way to show that a given model produced a given output. A commenter says it builds on VerInf, work done as a MARS fellow.

turn-051
00:16:31.960 --> 00:16:45.136
<v damra>Side by side, that's one lab saying the output may not be the model's best effort and another working on verifying where the output came from. Neither fixes the other. Proving which model ran doesn't tell you whether it tried.

turn-052
00:16:45.136 --> 00:17:06.616
<v lenar>Ricursive released Ripple-65B Preview, which it describes as the U.S. open-weight state of the art on agentic coding. It's a 65 billion parameter dense model, and the company claims it beats much larger open models. It had 310 points and 111 comments on Hacker News. The benchmark claims are Ricursive's own, and it's labeled a Preview.

turn-053
00:17:06.616 --> 00:17:22.984
<v damra>Commenters there say some teams are barred from using models from outside U.S. jurisdiction. That gives a U.S.-origin open model a market that has nothing to do with whether it tops a leaderboard. A model can be second-best and still be the only one a compliance office will approve.

turn-054
00:17:22.984 --> 00:17:44.464
<v lenar>Tencent also released Hy4, a 770 billion parameter model that SiliconANGLE says adds computer use for the first time. And Wired ran an interview with Hugging Face CEO Clément Delangue, who said his estimate of the probability of doom is, quote, 'close to zero,' and that, quote, 'Most of the risk is actually created by a couple of players building behind closed doors.'

turn-055
00:17:44.464 --> 00:17:58.912
<v damra>That's Delangue's view, from someone whose company hosts open models. It reads differently next to a closed lab's own report about a model underperforming, though that report is OpenAI disclosing the problem itself, and that counts for the closed labs too.

turn-056
00:17:58.912 --> 00:18:10.168
<v lenar>Next, a design from the AI Engineer talks. Emil Chetty, who leads data and AI product at Wise, argued that approving every agent step turns the customer into the agent's API.

turn-057
00:18:10.168 --> 00:18:11.824
<v damra>So what does he do instead?

turn-058
00:18:11.824 --> 00:18:43.552
<v lenar>Four stages. First, the agent explores with read-only access to gather context, like supplier details and amounts. Second, trusted, older software resolves discrepancies and produces a draft proposal. Third, the customer reviews a human-readable summary plus machine-readable bounds, authenticates, and approves the whole job once. Fourth, that approval becomes an immutable record, and a deterministic provisioning platform grants the agent only the intersection of the confirmed job, Wise's system permissions, and the user's current rights.

turn-059
00:18:43.552 --> 00:19:22.312
<v damra>The agent still picks its own steps; it just can't change the terms. If a quote expires, the system halts and goes back to the user for a new agreement. If a pod restarts, workers resume the same job instead of creating a new one. What I like is that downstream services enforce the record, so a prompt-injected agent can't redirect a payment, because the payment service checks the agreement and ignores what the agent says. And on the market side, The Information reports Palo Alto Networks is in talks to pay around 500 million dollars for Clutch Security, a three-year-old Israeli startup that tracks passwords used by AI agents. Talks only.

turn-060
00:19:22.312 --> 00:19:34.840
<v lenar>Chetty sets this against Anthropic's auto mode for coding agents, which cut friction with permission rules and sandboxes. For money, he says, the core question is whether the customer explicitly authorized those specific terms.

turn-061
00:19:34.840 --> 00:19:41.920
<v damra>Next, three AI Engineer talks that disagree about whether to build your own model or wait for the next frontier release.

turn-062
00:19:41.920 --> 00:20:04.360
<v lenar>Alex from Ramp says Ramp Labs post-trained a 35 billion parameter Qwen 3.5 model with Prime Intellect and beat Opus 4.6 on its finance tasks at lower latency. That's on Ramp's own tasks and self-reported. The same talk describes key-value cache 'cartridges' that compress recurring context up to 30 times, prepared offline for things like month-end reconciliation.

turn-063
00:20:04.360 --> 00:20:21.688
<v damra>And they ran 10,000 parallel agents on open-weight models to pen-test their own infrastructure. They found seven high-severity vulnerabilities, including a Lua script injection that needed analysis across components, at one-fifth the cost of frontier reasoning models and with comparable detection rates.

turn-064
00:20:21.688 --> 00:20:47.008
<v lenar>Blackstone's CTO John Stecher sits on the other side in the Norm AI talk. He says he often delays complex builds until models improve enough to lower the cost and effort, while prototyping quickly. He invokes the bitter lesson: systems that scale with compute eventually beat hand-built expert systems. Norm's Scott Worley says that when agent capabilities matured, much of the code they'd written around the model was obsolete and they discarded it.

turn-065
00:20:47.008 --> 00:21:15.208
<v damra>Norm throwing out its old code is the wait argument's best evidence. Ramp is the counterexample, though it's a company with a clear, repeated task and lots of volume. The third talk, from Nous Research's CEO, argues that labs will either trap customers or compete with them, so you should build on foundations you own. He's recommending Hermes Agent, which he describes as the most downloaded open-source agent, and Nous sells it, so he's an interested party.

turn-066
00:21:15.208 --> 00:21:37.648
<v lenar>The last item has a deadline attached. Google announced that schools restricting AI entirely will lose access to Google Search. The Guardian quotes Chris Phillips, Google's vice-president for education, telling the New York Times: 'search and AI are deeply integrated,' and so if a school has banned AI, students 'will not have access to search, and they'll see a message from their school that indicates it's not available.'

turn-067
00:21:37.648 --> 00:22:06.808
<v damra>So the message the student sees comes from the school, not from Google. TechCrunch reports that New York City mayor Zohran Mamdani and former FTC chair Lina Khan accused Google of acting as an unregulated monopoly, and the Guardian notes Google holds roughly 91 percent of global search share. No regulator has acted yet. But a Democratic Congress is about to take the committees, and this is an easy example to explain: one product conditioned on how a school treats a different product.

turn-068
00:22:06.808 --> 00:22:34.048
<v lenar>And one science item. Bloomberg reports that Isomorphic Labs, the DeepMind spinout, is in talks to raise about 5 billion dollars at roughly a 50 billion dollar valuation, after raising 2.1 billion in May. Google DeepMind posted about agentic cheminformatics on Antigravity, which it says compresses multi-day drug design workflows into an hour. And Nature has a piece on an automated biology lab where AI agents run DNA sequencing and other tasks.

turn-069
00:22:34.048 --> 00:22:58.408
<v damra>The DeepMind speedup is DeepMind's own claim and the Isomorphic round is talks, so both are claims for now. But days down to an hour is the kind of number investors price, and the Nature lab is the physical end of it: an agent that doesn't just propose the experiment but operates the sequencer. The detail that sticks with me from the whole day, though, is Lehane's two years. Those pledges were written for a window with a clock on it.

turn-070
00:22:58.408 --> 00:23:05.176
<v lenar>Monday's lame-duck session is the first place those pledges can turn into bill language or stay in a blueprint. I'm Lenar.
