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 the menu today is heavy on numbers that don't quite add up — funding rounds shrinking, chip stocks cratering, and a food-delivery CEO telling us the honest truth about AI agents that nobody in the AI business wants to say out loud. Pour the coffee, settle in, and let's have at it.
Here's where I want to start today, and it's not where the headlines wanted me to start. DoorDash's CEO got on an earnings call this week and said something that, if you squint, undercuts about a trillion dollars of industry hype. Tony Xu — the guy who runs DoorDash — was asked by an analyst how many orders are coming in through outside AI agents. You know, the whole dream where you tell your little assistant "order me dinner" and it goes off and does it. And Xu said, and I'm quoting him here, "AI agents are generating several hundred orders across the time that they've been up and running."
Several hundred. Let that sit for a second. DoorDash did just shy of one billion orders in the third quarter alone. Billion with a B. And the AI agents, the thing we've been told is going to eat the entire consumer internet, generated — across their entire existence — several hundred. Xu called it, in his words, "a very low use case." That, my friends, is what we call a marked understatement.
Now I want to be fair, because Xu was fair. DoorDash only rolled out its connector for third-party AI tools — the standardized hookup that lets an outside agent order through their platform — less than a month ago. It's in beta, a handful of early partners, mostly doing corporate lunch orders. So it's early. Pizza Hut's first online ordering experiment back in 1994 probably looked just as silly, and now more than half their orders come in online. Give it time, maybe.
But here's the line from Xu that I want every founder listening to tattoo on the inside of their eyelids. He said, "My take on all things technology is you never want to fight gravity. But at the same time, you don't want to be delusional." You never want to fight gravity, but you don't want to be delusional. That is the whole ballgame right there. And what Xu is telling you, with actual data from a billion orders, is that for a task as simple as ordering food, people just go directly to the app. They don't want a middleman robot. He sees agents being genuinely useful for what he called "long consideration purchases" — the big, slow decisions where an agent works in the background on your behalf. Not your Tuesday night burrito.
And notice — DoorDash's own in-house AI tool, the "Ask DoorDash" feature, that one's actually doing something. Baskets built five times faster, fifty percent higher basket value, sixty percent more unique items, and nearly half the orders landing at local spots the customer never tried before. So the AI that's woven into the product the customer already loves? Works great. The free-floating agent orchestrating from outside? Several hundred orders. That gap — between AI inside the thing people use and AI as a standalone butler — that's the story a lot of builders need to internalize before they raise their next round on an agent thesis.
Which, speaking of raising rounds, brings me to the money, and the money is getting nervous.
Let's talk about OpenAI, because the numbers moved and they moved the wrong direction. The Financial Times reported Friday that OpenAI lowered its funding target for its new round down to twenty billion dollars. Now, that might sound like a lot, and it is a lot, but here's the context. This same round was previously reported at thirty billion dollars, at a one-point-four trillion dollar valuation, and then at a one-point-two-five trillion valuation. So the target came down from thirty to twenty. And the Wall Street Journal added the detail that stings — the company has secured eleven billion so far, and its fundraising met with what the Journal called a "tepid" response from some large investors. The round's being led by a group of investors out of the United Arab Emirates, with a possible retail offering for wealthy individuals to help close the gap toward that twenty billion.
Tepid. From large investors. For OpenAI. Six months ago that sentence would've gotten you laughed out of the room. TechCrunch's framing is that the timing is just bad — Wall Street's gotten wary of heavy AI infrastructure spending, and there's been a broad sell-off in tech stocks. And we talked yesterday about that broader shift, the roughly seven hundred forty billion in AI capex that's suddenly spooking investors even though the AI itself works fine. This OpenAI number is the next data point in that same story. When the most prominent name in the business trims its round by a third and gets a lukewarm reception, that tells you the mood has genuinely turned.
And it's not just OpenAI. The same TechCrunch note mentions Anthropic's planned IPO could be facing challenges too, which lines up with what we covered — Blackstone warning Anthropic might land a below-investment-grade credit rating, a junk rating, after it goes public. A junk rating means higher borrowing costs, which means Anthropic leans harder on its stronger partners to back its debt.
And that right there is the thread connecting to maybe the wildest financial maneuver in the whole pile. There's a report — from The Information, corroborated by Bloomberg — that Nvidia is in talks to guarantee billions in debt financing for an AI cloud provider called Nscale, for a data center in West Virginia that Anthropic is the tenant of. Now follow the money here, because it's a circle. Anthropic has signed a forty-four-point-six billion dollar contract for just the first building of this campus — that's one building. The chips going in are Nvidia's next-generation Vera Rubin chips. And Nvidia is now considering co-signing Nscale's construction loans. Under what's called a residual value guarantee, if Nscale defaults on construction payments, Nvidia steps in, temporarily takes over the facility, and helps find a new buyer and tenant.
Think about the shape of that. The chipmaker is guaranteeing the loans of the cloud provider that's buying the chipmaker's chips to rent to the AI company that buys more of the chipmaker's chips than almost anybody. Nvidia has already extended something like a hundred seventy billion dollars in investments and credit support to these neoclouds and large customers. And why is Nvidia so eager to keep Nscale's project alive? An investor told The Information it's because Anthropic's the tenant — Nvidia wants to protect the customer.
Here's why a builder should care about this and not just gawk at it. These neoclouds — companies that rent out AI compute without a big cash-rich business underneath them — they're under real pressure. CoreWeave, Nebius, IREN — shares each down more than thirty percent in October. Nscale wants to go public at up to a thirty-five billion valuation while it's still struggling to get projects off the ground. It had about twenty-five thousand advanced chips running in August against a plan to deploy nearly five hundred thousand. Its West Virginia construction site has flooded nearby homes more than once. And Microsoft pulled out of a different Nscale site in England this month because the grid power got pushed into the 2030s. So the infrastructure layer underneath your AI products is being propped up by the chip vendor co-signing loans. That's not a healthy, self-sustaining market. That's a vendor-financed one. And vendor-financed booms have a way of reversing fast when the vendor blinks.
Now let's go from the circular financing to the thing that actually cracked open on Friday. Intel.
Intel's stock had one of its worst days ever — down around twenty-eight percent — after a weaker-than-expected outlook for the fourth quarter. And here's the kicker that makes it a political story too: the Trump administration had the government buy a ten percent stake in Intel over a year ago, and this one-day crash knocked about twelve billion dollars off the value of that government stake. Nvidia, which also took a big stake in Intel in 2025, slipped too. Much of the chip sector got caught in the selloff.
But I want you to hear the reason Intel gave, because it's genuinely important for anybody building hardware-dependent products. Intel's CEO Lip-Bu Tan said on the call — quoting — "Our customers began signaling a PC slowdown based on the memory prices and availability. Year-to-date system endpoint prices have increased significantly, placing additional pressure on already limited IT budgets." Translate that out of CEO-speak: the AI boom is gobbling up memory and components, that's driving up PC prices, and higher PC prices are killing PC sales. So the AI frenzy is literally cannibalizing the traditional business that pays Intel's bills. Intel's data center revenue grew seventy-eight percent year over year — great — but it's capacity-constrained, one new fab in Arizona running, the next not online until 2028.
And you see the same pattern over at Microsoft. We touched on their numbers — Azure up forty-eight percent, a monster. But in the fresh reporting this week, their Windows OEM revenue, the business of selling Windows licenses for new PCs, dropped twenty-seven percent. Their CFO Amy Hood said flat out they expect those trends to continue through the rest of the fiscal year. Same squeeze. The AI business is roaring, the legacy business that subsidizes it is sliding, and the AI business isn't big enough yet to stand on its own. That's the quiet tension underneath every one of these earnings reports, and it's why one bad outlook from Intel can take the whole sector down for a day.
Alright, let me shift from the balance sheets to the courtrooms and the regulators, because there's a cluster of accountability stories this week that founders should file away.
First, a judge let the NAACP's lawsuit against xAI move forward. This is about the Colossus data center outside Memphis, the one powering Grok. The Southern Environmental Law Center, suing on behalf of the NAACP, says xAI installed and operated at least thirty-five combustion turbines and other pollution sources without ever getting the required air permits — a Clean Air Act violation. xAI's defense was, frankly, creative. They argued the turbines are exempt because they're "mobile sources." Judge Debra M. Brown wasn't buying it. She ruled the turbines are not mobile sources under the state plan, that the plaintiffs have standing, and that these citizen suits aren't unconstitutional. Her line was clean: because citizen plaintiffs sue on their own behalf and seek penalties that redress harm to themselves, they're not exercising the federal government's executive power. A DOJ motion to dismiss got denied earlier this month too.
And here's the broader signal for anyone building anything that needs a data center: public opinion has turned hard against them. New York just passed a moratorium on new data centers. Other states and localities are weighing similar moves. The backlash even pushed Amazon and Microsoft to promise they'll stop using NDAs in their dealings with government agencies. If your product roadmap assumes you can just drop a gas-turbine-powered compute campus wherever the land is cheap and the questions are few — that era is closing. The community pushback is real, it's bipartisan, and it's now showing up in court.
Second accountability story, and this one's a security gut-punch for every developer listening. Anthropic's official VS Code extension — the one for Claude — was backdoored on Cursor's marketplace for nearly twelve hours. Dan Goodin at Ars reported it. A malicious version, 2.1.313, went out through Anthropic's verified, official account on Cursor, and because Cursor extensions auto-update by default, it hit as many as five thousand developers' machines. The backdoor tried to install a persistent SSH backdoor, open a reverse tunnel, and inventory your cloud credentials and API tokens so the attacker could prioritize which machines to hit next. Now, there were bugs in the malicious code that limited the damage in some cases — the SSH persistence often failed unless Cursor was running with admin privileges. But the security firm Phorian said it was still a significant threat.
And here's the part that should make you sit up. As of that reporting, neither Anthropic nor Cursor had explained how the attackers got control of the official account, and Anthropic hadn't publicly warned customers. Anthropic pushed a clean version, 2.1.314, several hours later — but as one commenter flagged, the clean version doesn't clean an already-infected machine. So if you installed or updated that extension on the Thursday in question, roughly between 5:27 AM Eastern and 4:35 PM, you should consider yourself compromised and go rotate your keys. This is the supply chain problem in its purest form. You can be a careful developer, trust a verified vendor's official channel, and still get owned because the trust itself got hijacked. Build your threat model accordingly — auto-update on dev tooling is convenience that doubles as attack surface.
Now let's dig into the deep dive, and it's the one I think matters most for how you think about building, careers, and the rules everyone's pretending to follow.
The provost of Dartmouth College, a man named Santiago Schnell, resigned this week — and the timing on this is almost too perfect. Back on August 9th, Schnell wrote an op-ed in the Washington Post titled "Universities are fighting AI cheating. But there's a deeper problem." This week he discovered the depth of that problem personally, when an advisory committee found he'd failed to disclose significant AI use in an academic paper and in that very op-ed, even though the publishers required disclosure. Ars's Nate Anderson had the full writeup, and I want to walk through it carefully because the surface story and the real story are different.
On the surface, this looks like a hypocrisy gotcha — the AI-cheating scold caught using AI. But read the actual details and it gets more interesting. The committee got access to Schnell's Claude enterprise account that ran through Dartmouth. And what did it show? It showed him using the model as an editing tool and a discussion partner. He'd run his papers through Claude and ask things like — and this is a direct quote from his prompt — "Imagine that you are a reviewer for this paper with expertise in nonlinear chemical dynamics and mathematical chemical kinetics. What would be your referee report for this paper?" That is not "write my paper for me." That's using the model as a tough pre-reviewer. Honestly, a lot of you listening have done exactly that this week.
Schnell himself described it as using AI "to refine a few arguments" and "check for grammatical errors." He told a journal editor he used it "to help refine sentences, improve clarity, and the expression of arguments." And he said he didn't disclose it to one journal because he "did not understand that this type of AI-assisted editing required disclosure." So — reasonable, right? Maybe.
But here's where the real story lives, and it's in a paper Schnell himself published in September called, fittingly, "Who Wrote This?" In it, he argues that the human-written-versus-AI-written binary is far too simple, that everything will become a mix of human judgment and AI work, and that what matters is the human author taking full responsibility. And then he lays out, in beautiful detail, exactly how he produced that very paper. Let me read you his own account, because it's remarkable. He writes:
"In September 2026, I asked OpenAI's GPT-5.6 Pro whether an AI-written essay could ever be identified with certainty... GPT-5.6 Pro connected the question to binary hypothesis testing and statistical decision theory, proposed early structures and titles, identified candidate literature, and generated portions of early drafts. I chose the problem, journal, audience, and eventual direction of the argument... I checked the cited literature, rejected or corrected claims that I judged inaccurate, corrected mathematical and conceptual errors, and reorganized the argument. In revising the manuscript, I rewrote most sections... I created the figures and made the final decisions about the mathematics, interpretation, sources, organization, and wording. I take full responsibility for the resulting manuscript."
Now that — that is a genuinely thoughtful disclosure. He's saying the AI generated drafts and structure, he directed and corrected and owned it, and he refuses to call the model a co-author because contribution and authorship are different roles. You could put that paragraph in a company handbook as the model for how to use AI honestly.
So here's the contradiction that sank him. He disclosed meticulously in the paper that was about disclosure — and then didn't disclose, even when required, on the other papers. And there's one exchange in the committee report that moves this from "confused about the rules" to something darker. Schnell uploaded a manuscript to Claude and prompted it: "Are there many signatures of AI writing in this article?" And then he submitted a revised version incorporating the changes Claude suggested. That's not refining arguments. That's asking the machine how to cover your tracks, then covering them.
But even that might not have been fatal. What was fatal — and an Ars commenter nailed this — is what Schnell did when the committee came knocking. He gave them his Dartmouth laptop but refused to hand over logs from his personal AI accounts related to the writings in question, and he wouldn't even meet with the committee. The report calls it "insufficient cooperation." And here's the thing — as provost, Schnell would normally be the guy in charge of academic misconduct investigations. A provost stonewalling a misconduct inquiry is like a police chief refusing to cooperate with an internal affairs review. Regardless of whether the underlying AI use was a real offense, the non-cooperation sealed it. How he handled the case became more damning than the case itself.
And here's why this is the deep dive and not a footnote. Dartmouth is barreling ahead with AI anyway. This is the school that literally coined the term "artificial intelligence" back in 1956. This same week it hosted a conference called "What Must Remain Human?" — with a student hackathon sponsored by Anthropic and speakers including Mira Murati and Yann LeCun. The president, Sian Leah Beilock, wrote a piece in The Atlantic in September called "Why I Want More AI at Dartmouth." So the institution's answer to "how do we handle AI" boils down to one word: disclosure. Beilock's statement on the resignation said leaders have to "model the transparency and human judgment that we expect from our students and scholars."
And the survey data from that conference tells you why disclosure is such a mess. Sixty-one percent of college students say AI helps explain difficult concepts. Half say they produce better work after talking through issues with AI. But seventy percent say take-home work has become unfair because of AI, and thirty-nine percent feel pressured to use it because their peers do. Among Dartmouth's own students, ninety-two percent fear AI can hurt their writing skills. Everybody's using it, everybody suspects everybody else is using more, and nobody agrees on what has to be said out loud.
Here's the takeaway for founders and builders, because this is your problem too, right now, in your company. The entire governance model everyone is landing on — in universities, in enterprises, in your own shop — is disclosure. Tell people when you used AI. But Schnell's case exposes how slippery that is. What counts as "AI use" that needs disclosing? Grammar check? No. Generating full draft sections and structure? Clearly yes. Asking the model to help you hide that you used the model? That's the line into misconduct. And the uncomfortable truth is that nobody has drawn those lines cleanly, which means right now the enforcement is basically a vibes check — the Ars commenter's phrase — until someone decides to make an example of you. If you run a team, don't leave this to vibes. Write down what your disclosure standard actually is, concretely, with examples. Because the cost of ambiguity isn't theoretical anymore. It cost a man who literally studies this his job.
Alright, let me bring it back down to earth with a couple of shorter ones before we close, because not everything today is doom and circular finance.
Salesforce keeps buying. This week it's a reported three hundred seventy-five million dollar deal for a defense-tech startup called Rune Technologies — AI software that helps the US military predict supply needs and manage battlefield logistics. This feeds Salesforce's defense unit, Missionforce, which launched last year and has grown eighty percent year over year, partly on the back of a ten-year, five-point-six billion dollar Army contract. And it's part of a serious acquisition spree — Salesforce also agreed to buy the AI agent platform Fin, formerly Intercom, for three-point-six billion, bought Qualified for one-point-two billion, and grabbed Listen Labs after that startup walked away from its own funding round. Rune, for its part, was last valued at a hundred twenty-seven million and had raised thirty million from backers including Andreessen Horowitz and IQT, the CIA-backed venture firm formerly known as In-Q-Tel. The pattern to watch: the CRM giant is quietly becoming a defense software player, and enterprise AI consolidation is happening at speed.
And a quick note from the other end of that spectrum — a startup called Ollie, a privacy-focused AI assistant, is shutting down on November 6th. This was one of the first mainstream apps to actually achieve SOC 2 privacy compliance. It raised a seven-and-a-half million dollar seed from Khosla Ventures just back in September. The co-founder Bill Lennon said it straight: "We got close, but ultimately couldn't keep up. With Muse, Instinct, and others entering the space, it became harder to stand out, grow, and raise the capital we needed to succeed." The lesson there is bleak — privacy as a product differentiator still doesn't pay the bills in this industry. A worthy mission, a real compliance achievement, and it still couldn't clear the fundraising bar in a market where everyone's piling into the same two or three names. Which, if you connect it back to that OpenAI round getting a tepid reception — even the giants are feeling the squeeze, so the little privacy-first player never had much runway.
And one from the meta-layer, OpenAI had a run of policy posts this week worth a one-line flag each for the builders: they published a defense of fair use for AI training, sharing excerpts from court briefs by creators and academics on why fair use supports AI training; they put out a community safety report with new protections for teens in ChatGPT; and they committed to honoring a shared robots-dot-txt signal to let website operators express their AI training preferences. That last one matters most for anyone who runs a site — there's now a clearer, standardized way to tell participating AI providers whether they can train on your content. Whether they all honor it is the open question, but the plumbing is getting built.
Let me leave you with the throughline, because I think it's real and not forced. The connecting tissue in today's menu is a gap between what's working and what's getting funded, trusted, or governed. DoorDash shows AI agents barely used while the hype says they're everywhere. OpenAI's round shrinks while the models keep improving. Nvidia's co-signing loans to keep the infrastructure layer standing. And a provost who studies AI honesty for a living couldn't navigate the disclosure rules his own institution is betting its future on. The technology is further along than the structures we've built to pay for it, trust it, and account for it. That's not a reason to bail — it's a reason to build the boring stuff: the disclosure standards, the key rotation, the honest read on whether your agent is actually getting used. The companies that survive the next year won't be the ones with the flashiest demo. They'll be the ones who, like Tony Xu said, don't fight gravity but don't get delusional either.
That's the show, kiddos. I'm Tony DeLuca, this has been Barely Possible, and remember — when somebody's guaranteeing their own customer's loans, keep one hand on your wallet. Take care of each other out there.