Four billion dollars in compute contracts signed before lunch, and neither company signing them is one you've heard of. This is AI Daily Briefing. Today: the week compute-deal numbers stopped being theoretical and started showing up as receipts — plus Netflix spends half a billion on an Affleck AI startup with nothing on record about what it actually does. Bill, start me on IREN. Tap follow so the next episode finds you. William Foxley, writing in Blockspace:
IREN (NASDAQ: IREN) said Monday that it had signed $2.8 billion of new multi-year AI cloud services contracts with leading AI developers. The data center and bitcoin mining operator lifted its year-end 2026 AI cloud annualized run-rate revenue target from $3.7 billion to more than $4 billion.
IREN — $2.8 billion in new multi-year AI cloud contracts, ARR target bumped from $3.7 to over $4 billion, stock up 15% Monday. Two months ago most of our listeners couldn't have picked this company out of a lineup. And here's the number I actually like — customer prepayments covering about 45% of the GPU capex. Somebody's putting cash down before the racks are lit. Cash up front says a lot more than a press release. This is a bitcoin miner. Danny Roberts says IREN went from roughly 3 megawatts of self-built AI cloud to 480 in twelve months. Capacity New York and other tight markets couldn't take is landing with companies you've never heard of. But watch that word 'annualized' — run-rate isn't cash in the door. They're at 85% of a target that's a projection of a projection. Four-year weighted terms means most of this dollar figure lands slowly, not this quarter. IT Brief New Zealand, with Sean Mitchell:
Sharon AI has signed a five-year cloud computing services agreement worth USD $1.32 billion with an unnamed global artificial intelligence laboratory. Services will be delivered from data centre infrastructure in New Zealand. Revenue is expected to begin progressively during the first and second quarters of 2027.
Right after IREN's $2.8 billion, here's another one: Sharon AI, five-year deal, $1.32 billion, landing in New Zealand. Not where anyone was drawing the compute map six months ago. And it's one customer. An unnamed global AI lab, the whole $1.32 billion, single counterparty. So the risk is all in one place — one very large egg in one basket. The anonymity matters. If you're a frontier lab or a hyperscaler, you usually want your logo on this. Staying quiet costs you the PR win — so whatever the reason is, they value it more than the headline. Revenue doesn't even start trickling in until Q1 or Q2 of 2027. So the headline number is real, but the cash starts months from now and is back-loaded across five years. Announce now, collect later. And New Zealand is hosting frontier-scale AI for the first time — 62,000 Nvidia GPUs by mid-2027, up from 55,000 in June. The demand New York and other markets turned away has to go somewhere, and today somewhere is Auckland-adjacent. ChosunBiz, with Jeong Du-yong:
The negotiations are in the early stages, and there is a possibility they may not lead to a final contract. The fact that Meta has not previously operated a business selling computing resources to external parties is another variable in the negotiations. Meta did not respond to related inquiries, and Anthropic declined to comment.
So after IREN's $2.8 billion and Sharon AI's $1.32 billion, here comes a $10 billion Meta-Anthropic number — except read the first line of the article. Early stage, may not lead to a contract. Right, that's the fine print sitting right next to a ten-billion-dollar headline. But Anthropic even being in the room asking Meta for capacity tells you how bad the GPU crunch has gotten. This is a lab with its own campus and a services arm, and it still needs to rent from Meta of all people. Anthropic and OpenAI love to talk about how different they are. They're scrambling for the same GPUs. And Meta's never sold compute to outsiders. Zuckerberg floated this at the May shareholder meeting — companies calling every week to buy surplus capacity. Now Meta wants a CoreWeave-like business to loosen the ad-revenue dependence. Which means Meta could end up underwriting a competitor's frontier models with its spare racks. Everybody's a landlord now. A $10 billion lease that may never close is the disclaimer I want stapled to every announcement this week. Announcing a number and stress-testing the contract are different sports. Here's Anthony Ha at TechCrunch:
In a new regulatory filing, Netflix revealed that it paid $587 million in cash for InterPositive, a startup co-founded by actor and director Ben Affleck. The streaming company announced the acquisition in March, with a statement from Affleck saying he wanted to “protect the power of human creativity.”
Okay, after the billion-dollar compute parade, here's the rounding error by comparison — Netflix paid $587 million in cash for Ben Affleck's InterPositive. And the number matters, because that's more than a lot of infra companies raised all year for a creative-workflow tool. And the $587 million only shows up now because a regulatory filing forced it out — back in March, when they announced the deal, they kept the terms quiet. Bloomberg had guessed up to $600 million. Affleck's pitch was he wanted to "protect the power of human creativity." Half a billion dollars to protect creativity by automating the reshoots. Sure. What I don't have is any technical report on what InterPositive actually does at production scale. Fixing missing shots, background replacement, lighting — those are demo-reel bullets. They don't tell me the spec sheet. Netflix says 300 titles already used generative AI, but not which ones or how much. And here's why this one is hard for me — this is the one AI category where I can't hide behind step-seven-of-ten failure rates. The stress test is box office. Does the finished film survive an audience, not a benchmark. TechTimes, with Scott McCain:
The result is 96.7% accuracy on MNIST and 61.7% on CIFAR-10 — without backpropagation, without random feedback matrices, and without violating the biological constraint that individual neurons can only excite or only inhibit, never both. The gap to the nearest competing backpropagation-free method is 7.4 percentage points on CIFAR-10.
Okay, let's downshift from ten-billion-dollar leases to a paper that cost nobody ten billion dollars and I actually find it more interesting. Sakana AI, in Tokyo, posted this to arXiv on the 17th — training a neural net without backpropagation. The honest part is right there in the framing. They call it the biology tax. 96.7% on MNIST, 61.7% on CIFAR-10 — and they give you the gap to the nearest backprop-free method: 7.4 points. Right — they published the number that makes them look worse. That's what makes me take it seriously. The whole point is measuring, in accuracy, what biologically plausible learning costs you today. And there's actually a paper attached — Diffusing Blame, on arXiv. After the week we've had, a release where the technical report is the release feels almost quaint. The backprop problem matters because neuroscientists have said since 1989 that the brain can't do the math backprop needs — Crick himself said it. Nearly forty years later, somebody's finally putting a price tag on the alternative. Before I get too excited: CIFAR-10 is still a toy by 2026 standards. 61.7% is a lab curiosity, not a phone camera. But they got it working with convolutional nets and reinforcement learning, and that's the first time anyone's shown that. So, real step — just a small one on a very long ladder. If you're following the AI power players we cover here, try Musk v Altman Daily — a daily court-watch on Elon Musk’s trial against Sam Altman, OpenAI, and Microsoft, tracking testimony, exhibits, and the AGI governance fight. Find it wherever you listen to podcasts.
What we’re watching next: IREN says it’s scheduled to deliver 480 megawatts of AI cloud capacity in 2026, with 1.2 gigawatts targeted for 2027. Sharon AI expects revenue from its New Zealand cloud agreement to start coming in during the first and second quarters of 2027.
Links to every story we mentioned are in the show notes, so if one caught your ear, you can dig in there. That’s AI Daily Briefing for today. This is a Lantern Podcast.