TensorWave just raised $350 million to build AI clouds on AMD — and that one word, AMD, is the most interesting thing I've heard in tech all week. This is the AI Daily Briefing. Today: the compute race fans out from those AMD clouds to Jamnagar and Harlow, plus a New York disclosure law that just went live. And for once the rundown's lighter on dollars and heavier on real numbers. I can work with that. Let's start with TensorWave, because the silicon bet is the story everyone's going to cite and almost nobody's going to actually read. $350 million is below the line where I'd flag it on size alone. But all-AMD? That's the first funding story this week with an actual silicon bet attached. Every other cloud buildout this week has been an Nvidia story. This one is a structural bet against Jensen owning the inference layer — the layer that actually decides what it costs to run anything. Meta-RIL next — 168 megawatts in Jamnagar, with Reliance attached and a real site. And notice who Meta anchored to — Reliance, not a government rack. In India, with the data-residency politics there, who controls fine-tuning inside that facility is exactly the thing 'sovereign' papers over. Right, what did Meta promise Reliance on residency and latency to get that signature? That's the term sheet I'd want, and it's the one we never see. Nebius Harlow has edges on it now — a ten-year agreement, 22 megawatts, the Kao Data campus. Last week it was a London postcode on Jensen's roadmap; today it's a contract. Quick one on DiffusionGemma — it ships on a Gemma 4 backbone, and the developer guide actually describes serving behavior. The word 'compute-bound' is in the lede. And Hochul's AI ad disclosure law officially went live yesterday. So she's got two disclosure regimes running at once now — and neither touches the inference stack. The week opened on military memos and closed on megawatts and enforcement clocks. That infrastructure beat I keep being told is sleepy has been the whole show three days running. Here's Ventureburn:
TensorWave has successfully closed a $350 million Series B funding round. The investment, which values the three-year-old startup at $1.55 billion, signals a major market bet on AMD-powered infrastructure as a viable alternative for the world’s most demanding AI workloads.
Buried in this one is the detail that actually matters. TensorWave is all-AMD — Instinct accelerators, the ROCm stack — and it rejected H100 and Blackwell entirely. Every other cloud buildout this week has been an Nvidia story. A $350 million bet on AMD's inference stack is a very specific wager against Jensen owning the layer that does the actual work. And here's the part nobody puts in the press release. $350 million is below the line where I'd raise an eyebrow on principle. What matters is whether ROCm survives the same production workloads CUDA has spent a decade hardening. CUDA's moat comes from ten years of people debugging it at 2 a.m., not from the silicon alone. AMD's software stack has to eat that — and the round size doesn't tell you whether it has. The other tell: AMD Ventures co-led this. So the chip supplier is funding its own biggest customer off its own balance sheet — exactly the playbook Nvidia ran for years to lock in share. Aabhas Sharma, writing in The Times of India:
NEW DELHI: Meta has signed its first artificial intelligence data centre deal in India, partnering with Reliance Industries on a 168-megawatt facility in Jamnagar, Gujarat, as technology companies race to secure the computing infrastructure needed to power AI applications.
168 megawatts in Jamnagar, with Reliance on the other side and a two-year operational timeline. After a week of gigawatt press briefings that aren't drawing power, this is the first one I'd actually put in the 'probably runs inference' column. And read the division of labor, Bill. Reliance handles design, construction, connectivity, operations. Meta pays for energy and water and leases the capacity. So who controls fine-tuning and inference inside that facility? India's data-residency politics are even heavier than France's, and Meta's anchoring to Reliance — not a government rack. Right, and that's the term I want — what residency and latency commitments did Meta make to get Reliance to sign? Renewable power, desalinated seawater for cooling, that's the headline-friendly part. The contract terms are where the real deal lives. Remember, Meta put $5.7 billion into Jio Platforms back in 2020, so this deal has history. The data-center piece is the six-year-old relationship growing a power bill. Here's Ian Ballantyne, Omar Sanseviero at Google Developers Blog:
Compute-bound parallel generation: Bypasses memory-bandwidth limitations by shifting the bottleneck to compute, delivering up to 4x faster token generation on GPUs (up to 700+ tokens per second on NVIDIA GeForce RTX 5090 and 1000+ tokens per second on a single NVIDIA H100).
Finally — a model release I can actually read. DiffusionGemma ships with a guide that tells you how it serves, and the word in the lede is 'compute-bound.' That's the first sentence worth reading. After a week of megawatts and funding rounds, somebody attached an actual technical document to a launch. I almost don't know what to do with myself. Here's why compute-bound matters: autoregressive models choke on memory bandwidth because they're reloading weights for every token. Move the bottleneck to compute, and Google is claiming 1000-plus tokens a second on a single H100. 700-plus on a 5090 — that's a consumer card. And it's a 26-billion-parameter Mixture of Experts that only fires 3.8 billion parameters at inference. Gemma 4 backbone, quantized inside 18 gigs of VRAM. That 18-gig number is the tell. It means a developer on a single mid-tier accelerator can run this — the cost surface drops out the bottom. That's the line in the press release nobody usually gives you, because it's the one that survives contact with a real latency budget. Bidirectional attention doing self-correction across the whole block too — which is genuinely a different shape than the one-token-at-a-time loop everyone's been optimizing. I'll wait to see the failure modes, but the document at least tells you where to look. It says 'experimental,' which I read as Google telling you the error rate isn't pinned down yet. But they told you. That alone puts this ahead of half the agentic demos I've watched this month. Techerati, with Rebecca Uffindell:
Nebius has signed a 10-year agreement to deploy 22MW of AI infrastructure at Kao Data’s Harlow campus, one of the largest AI cloud deployments announced in the UK to date. The agreement supports AI cloud services, production-scale inference and academic research, while offering a glimpse into how the UK’s AI ambitions are beginning to appear in the form of deployed infrastructure rather than policy alone.
Okay, this is the one I'd put in the runs-inference column: Nebius, 22 megawatts, a ten-year deal at Kao Data's Harlow campus. With the site and counterparty actually named, that feels refreshing after this week. And this is the one I flagged Monday as Nebius reselling Jensen's roadmap with a London postcode. Now the deal's got edges — 22 megawatts, ten years, part of a £1.7 billion UK commitment. But the ten-year term opens up the thing I want answered: is the facility locked to Harlow for a decade, or is Nebius locking in Nvidia architecture for a decade? Those are very different bets. Right — a long-term building lease is normal. If the silicon roadmap is fixed for ten years, that's the line I'd want clarified. Twenty-two megawatts of managed inference is a serious commitment to whatever's in the racks. And it answers the UK legibility question we were chewing on Monday. There's secured capacity here, with a contract behind it. WBNG, with Jolie Jenkins:
NEW YORK (WBNG) -- Gov. Kathy Hochul announced Tuesday that the first-in-the-nation law to boost AI transparency in advertising in the film and television industry is now in effect. Senate Bill S8420A, signed in December 2025, was introduced in response to the growing use of AI-generated performers across all forms of media, including social media and digital advertising. It requires advertisement producers and creators to identify if their content includes AI-generated synthetic performers, which are digitally created media that appear as a real person.
So the synthetic-performer disclosure law we had in the rundown yesterday as pending — it's live as of June 10. S8420A, signed back in December, now has an enforcement clock. And here's what gets me: Hochul now has two active disclosure regimes running at once, and neither one touches the inference stack. You have to label the synthetic performer in the ad — fine. Nobody has to say a word about who controls the model that generated it. What I keep noticing across this whole week's regulatory pile is that today's the first one to actually cross the 'in effect' line. Everything else has been signed-but-pending. Which makes Connecticut's hiring liability date in October feel a lot more concrete all of a sudden. We've now got a real before-and-after example of what 'in effect' looks like — and I'd bet most founders building agentic hiring pipelines haven't priced that compliance date at all. Right, the law reaches the label on the ad. It doesn't reach the layer that actually decides anything. Label the output, ignore the stack. Got thoughts on today's briefing, a story idea we should track, or a correction? Send us a note at aidailybriefing at lantern podcasts dot com. We read the inbox, and we appreciate hearing from you.
You'll find links to every story we covered today in the show notes if you want to dig deeper or pass along the pieces that stood out.
That's AI Daily Briefing for today. This is a Lantern Podcast.