Japan more or less wrote the playbook on modern manufacturing: the assembly lines, the robots, the relentless efficiency. Now it wants to do the same for artificial intelligence, and it's building a very large computer to get there.

What it is

Nvidia has announced that it's working with Japan's Noetra Corp. to build a 140-megawatt "AI factory," a facility packed with 27,500 Rubin GPUs and 13,750 Vera CPUs. (Rubin is Nvidia's next generation of AI chips, following its current Blackwell line, and a GPU is the processor that does the heavy lifting in AI training.) For scale, 140 megawatts is roughly the power draw of a small city's worth of homes.

The hardware is built from 382 of Nvidia's Vera Rubin NVL72 racks, each holding 72 GPUs and 36 CPUs, wired together with Nvidia's Spectrum-X Ethernet networking. This is the compute foundation for FRONTia, a state-funded Japanese program to train open, multimodal AI models for robotics, digital twins, and industrial automation. Multimodal simply means the models can handle more than text, so images, video, and audio too. Crucially, the pretrained models will be shared broadly with domestic developers rather than locked away.

"Japan invented modern manufacturing. Now, it is building the AI factories that will power the next industrial revolution," said Nvidia CEO Jensen Huang in the announcement.

Who is behind it

Noetra is a brand-new consortium founded by SoftBank, Sony, NEC, and Honda, with backing from 44 companies and organizations. Together with the national research institute AIST, it won a public tender on June 30 to run FRONTia from fiscal 2026 through fiscal 2030. The funding is substantial: roughly $2.4 billion in the first year, and up to about $6.1 billion over five years, according to Asia Times.

One honest caveat: that full amount is not guaranteed. Funding beyond the first two years is subject to annual stage-gate reviews, meaning the government can pull back if progress disappoints.

Why it matters

Plenty of big AI computers already exist. What makes this one different is that it's national infrastructure, tendered by the state, rather than a private corporate cluster or a one-off scientific supercomputer. It follows SoftBank's Blackwell-based supercomputer from 2024 and FugakuNEXT, a $740 million system due around 2030, but it's the first pitched explicitly as shared public groundwork.

The strategy behind it is specific. Japan's AI Robotics Strategy, released in March, aims to capture more than 30% of the global AI robotics market by 2040, an opportunity the government pegs at $133 billion. Rather than let each company train its own foundation models from scratch, the idea is to build the models once, with public money, and let domestic developers build on top.

The price tag is not officially disclosed, but the math is eye-catching. Vera Rubin NVL72 systems are currently quoted at $5 million to $7 million each, which puts the rack hardware alone somewhere between $1.9 billion and $2.7 billion. Morgan Stanley estimates Nvidia will charge $55,000 per Rubin GPU in volume, pricing the raw GPU silicon at roughly $1.5 billion before you add memory, networking, and cooling. Treat those as analyst estimates, not confirmed invoices.

What's next

No deployment timeline was given, and there's a practical reason for caution: Rubin racks aren't expected to reach volume production until the second half of this year. Nvidia said the facility will support trillion-parameter model training "as the AI factory expands," which hints at a phased build rather than a single switch-on.

Noetra's roadmap is staged too. It targets a reasoning foundation model in fiscal 2026, an omni-modal model that handles text, images, video, and audio by fiscal 2028, and "real-world native AI" with spatial awareness by fiscal 2030.

The interesting question isn't whether Japan can buy enough GPUs; with this budget, it clearly can. It's whether pooling national resources into shared, open models produces something more useful than the private clusters everyone else is racing to build. If it works, expect other governments to reach for the same blueprint.

Building a clever AI model is hard. Getting a giant, non-tech company to actually use it well might be even harder, and possibly more lucrative. That is the bet behind Ode with Anthropic, a $1.5 billion company launched in May by the AI lab Anthropic alongside Blackstone, Hellman & Friedman, Goldman Sachs, and others.

The idea is simple to state and tricky to pull off. Frontier labs, the companies building the most advanced AI systems, are realizing that shipping a smarter model isn't enough to win big business customers. Someone has to walk into the office and rewire how the company actually works. Ode wants to be that someone.

What Ode actually is

Ode began as an observation. Blackstone, the private equity giant, kept hiring consultants and small AI shops to roll out AI across the companies it owns. One boutique, a startup called Fractional AI, stood out. So the joint venture bought it. Fractional, which ended an 11-month partnership with OpenAI on acquisition, is now the foundation of Ode, described by its leaders as a "scaled boutique" AI services firm.

Today it employs around 100 engineers and works closely with Anthropic's applied AI team to find where the technology can help a given business, then build custom systems around it. Ode runs on a "Claude-first" principle, meaning it reaches for Anthropic's tools (Claude is Anthropic's AI assistant) when it can, but it will use rival products if a job calls for it.

CEO Chris Taylor, a Fractional co-founder, is not shy about ambitions. "It's pretty easy to imagine this as a trillion-dollar company someday if we execute well," he told TechCrunch. That is a projection, not a promise, and worth reading as one.

Why implementation, not models

Ode's pitch is that model choice matters less than people think. "I think model selection matters, but it's not where the majority of calories are spent," said chief technologist and co-founder Eddie Siegel. He compares picking a model to picking a programming language: important, but not the thing that defines whether a company transforms itself.

The harder work is taking what Taylor calls "this magic, hallucinating ingredient" (a nod to AI's habit of confidently making things up) and using it to rebuild core business processes without breaking them. That takes talent most companies don't have in-house.

Ode describes its team as elite generalist engineers, over half of them former startup founders. One Blackstone executive called them "special forces" rather than an army of forward-deployed engineers, the on-site staff labs send to embed with clients. Ode's ideal customer is a company whose CEO personally buys in, treating the project as a top-one-or-two priority.

The catch

The plan has an obvious tension. Ode wants to scale fast, internationally even, while keeping its boutique quality. But its secret sauce is exactly the kind of engineer who is scarce and expensive: entrepreneurial, systems-minded, with real product judgment. Can you train enough of those people to meet demand? That is the open question hanging over the whole venture.

Competition is stacking up too. Ode will square off against OpenAI's own version, The Deployment Company, plus consulting heavyweights like Deloitte and Accenture, which have built their own deployment teams. Demand for these teams, everyone involved agrees, far outstrips supply.

Siegel isn't fretting about the talent pool. "It has never been an easier time to become an entrepreneur," he said, arguing that owning a problem end-to-end teaches skills you can't get solving narrow tasks.

What's next

The private equity backers will steer their portfolio companies toward Ode as ready-made customers, though Ode says it won't limit itself to them. Anthropic's internal team, meanwhile, will keep handling its own strategic deployments.

The bigger takeaway is where the industry thinks value is moving. If Ode and its backers are right, the next phase of the AI race won't be won purely on who has the smartest model. It will be won by whoever can actually get that model working inside the world's largest, least tech-native companies. Whether enough "grown-up" engineers show up to do it is the part nobody can yet answer.

For weeks this summer, the AI world was glued to Anthropic's newest frontier models and the Washington fight over who gets to use them. Meanwhile, developers quietly went about their business. They weren't waiting for permission from OpenAI or Anthropic. They were building on something cheaper.

What the numbers show

This spring, Chinese open-weight models accounted for 41% of downloads on Hugging Face, edging past US models. ("Open-weight" means the model's trained parameters are published, so anyone can download, run, and tweak it rather than renting access through a company's API.) On OpenRouter, a service that routes requests to different models, the six most popular are all open models from Chinese firms including Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai. Anthropic's Claude Opus 4.7 sits in seventh.

Vercel, a platform for deploying web apps, reports a similar split. Open models handled nearly a third of AI requests there in June, absorbing the high-volume, everyday workloads while pricier closed models sit in the premium tier.

A fair caveat: these platforms only capture one slice of the market. They miss sessions hosted directly by the big labs, which likely make up the bulk of OpenAI and Anthropic's usage. Still, the trend is hard to ignore.

Rent or own

The question this raises is uncomfortable for frontier labs: how much do the smartest models matter if most production AI runs on cheaper, customizable ones?

Hugging Face CEO Clem Delangue thinks the top-tier models may end up reserved for experiments and a handful of high-value tasks, while the day-to-day work runs on open source or companies' own private models. Hugging Face, for the record, hosts and helps deploy open models, so it has a stake in this view.

His pitch is about ownership. "If you're an AI company or a technology company, you don't want to outsource your core capabilities to another company, to a black box API that you don't control," Delangue said. He says a new repository appears on the platform every seven seconds, and it now hosts almost three million public models and a million public datasets. Half of the Fortune 500, he claims, use Hugging Face to deploy private or open models.

Microsoft CEO Satya Nadella has made a related argument against "single provider lock-in," warning that if learning only flows one way, value drifts toward whoever owns the infrastructure rather than the people creating the knowledge.

The safety fight

The steady drip of capable Chinese releases keeps feeding this shift. The latest is Z.ai's GLM-5.2, an open model tuned for agentic coding that competes with Anthropic's models on spotting security vulnerabilities.

Not everyone is thrilled. Anthropic CEO Dario Amodei argues that releasing powerful open weights is risky because once they're out, they can't be recalled. Critics warn bad actors could use them for disinformation or worse.

Delangue flips the framing. "The biggest risk in AI is concentration of power," he said. In his view, transparency lets defenders patch known weaknesses, while locking models away just creates an "asymmetry of power." He also notes that closed models aren't airtight, since guardrails get bypassed and weights get stolen.

What's next

The interesting shift here is that "winning" AI may no longer mean having the single smartest model. It may mean having the cheapest, most controllable one for the job in front of you. If Delangue is right, the frontier becomes a research lab and a premium option rather than the main event, while most real-world AI runs quietly on models companies can own and shape.

Whether regulators, and the frontier labs themselves, are comfortable with that future is the fight worth watching next.

Somewhere in Richland Parish, Louisiana, Meta is building a computer so hungry it needs its own gas plants. On July 13, the company confirmed it will expand its Hyperion campus to 5 gigawatts (GW) of compute capacity, up from an initial 2 GW. For scale, a gigawatt is roughly the output of a large power station, so five of them could run a mid-sized city.

That expansion pushes Meta's planned investment in the region past $50 billion. It is a steep climb from the $10 billion, 4-million-square-foot project Meta first unveiled in December 2024.

What Hyperion actually is

Hyperion is Meta's AI supercluster, a giant network of connected servers built to train and run the company's future AI models. CEO Mark Zuckerberg has tied it directly to Meta Superintelligence Labs, the firm's AI division. The 5GW target is not brand new; Zuckerberg floated that figure back in July 2025. What is new is the formal price tag and fresh numbers on jobs, contracts, and public spending.

The financing is worth a glance too. In October 2025, Meta and Blue Owl Capital formed a joint venture valuing the buildings and infrastructure at about $27 billion, with Blue Owl holding roughly 80% and Meta keeping 20% and leasing the finished facilities. Monday's announcement did not explain how the expansion changes that arrangement.

Why the charm offensive

Much of Meta's message was aimed at locals, and for good reason. Data centers this large strain electricity grids and water supplies, and nearby residents worry about their bills. So Meta leaned into the economic upside. Local Louisiana businesses have received more than $1.6 billion in contracts since construction began. Teacher bonuses in Richland Parish rose from $10,000 last year to more than $50,000 this year, funded by tax revenue linked to the campus.

Meta also pledged over $1 billion for local infrastructure, including roads, water, and wastewater systems. Its deal with utility Entergy Louisiana includes natural-gas plants supplying more than 5.2 GW, plus support for up to 2.5 GW of new solar. Entergy says Meta's payments could save other customers around $2 billion over 20 years. That is a projection, not a guarantee, so treat it as a hopeful forecast rather than a done deal.

Who pays, who benefits

Meta is also getting generous public help. In late 2024, Governor Jeff Landry signed a 20-year sales-tax exemption for data centers built before 2029, an explicit move to lure Meta. The company also stands to benefit from the state's Quality Jobs program and a payment-in-lieu-of-taxes deal that could trim its property-tax bill if it hits investment and employment targets.

Hyperion is just one node in a much bigger spend. Meta is forecast to lay out up to $145 billion in capital expenditure in 2026, mostly on AI infrastructure, as demand for AI compute keeps outrunning supply. To help fund it, the company has said it will cut 8,000 jobs. The timing is striking: pouring tens of billions into silicon and gas turbines while trimming headcount.

What's next

The announcement landed after what Meta calls its strongest market week since early 2024, driven by new AI model releases. That momentum helps explain the confidence to keep scaling. The open questions are less about ambition and more about follow-through. Will the promised customer savings and infrastructure upgrades materialize, or stay on the projection slide? And can a rural parish absorb 5 GW of demand without residents feeling the squeeze? Louisiana just became one of the clearest test cases for whether the AI buildout pays off locally, or simply plugs in and pulls hard.

The plumbing of AI took center stage this week: the memory that feeds hungry chips, the chips themselves, and the tools that put models within reach. We also found time for robots learning to picture the future. Let's dig in.

Teaching a machine to grab a red cube is harder than it sounds, mostly because the robot has no idea whether it succeeded. LeRobot Teaches Robots to Imagine covers Hugging Face's v0.6.0 update, which adds policies that picture what happens next, models that grade success, and a way to turn failures into training data. The theme is closing the loop so robots can learn from their own mistakes.

AI chips have a feeding problem: processors run so fast that memory can't shovel data quickly enough. Intel's XBM Patent Rethinks AI Memory looks at a fresh idea to ditch HBM's costly silicon interposer in favor of a chiplet-native memory stack. It's a clever swing at the memory wall, though for now it's just a patent.

The humble CPU rarely gets a mention in the AI story, and Nvidia wants to fix that. Nvidia Bets on Speed Over Core Count unpacks its pitch for the upcoming Arm-based Vera chip, arguing that AI agents don't need more cores, just faster ones.

If you've ever fought to run an open AI model on your own laptop, this one's for you. Ollama Raises $65M as Users Near 9M tracks the tool that makes local models painless, now backed by fresh funding and sitting inside 85% of the Fortune 500.

And the big number of the week: a South Korean company that nearly went bankrupt in 2001 just made history. SK hynix's Record $26.5B US IPO details the largest foreign IPO in US history, with demand more than seven times the shares on offer, and every dollar headed toward making more chips.

Next week we'll keep watching whether these bets on faster, cheaper AI hardware start showing up in the tools you actually use. See you then.

On July 10, a South Korean company that nearly went bankrupt in 2001 completed the largest foreign IPO in US history. SK hynix raised $26.5 billion in its Nasdaq debut, selling 177.9 million American depositary receipts (ADRs), a US-listed stand-in for a foreign share, at $149 each. Each ADR represents a tenth of a Seoul-listed share. Demand was more than seven times what was on offer, drawing over 500 investment firms, according to the Financial Times. Not bad for a business that posted a heavy annual loss just two years ago.

The stock is trading temporarily under the ticker SKHYV before regular-way trading begins as SKHY on Monday, July 13. So what turned this company into one of the hottest names on the market? Three letters: HBM.

What SK hynix actually makes

SK hynix is the world's leading producer of high-bandwidth memory, or HBM. That is a mouthful, so here is the plain version: HBM is computer memory chips stacked vertically like a tiny apartment tower, which lets data move much faster than traditional flat memory layouts. That speed matters enormously for AI accelerators, the specialized chips that train and run large AI models. In short, HBM has become critical plumbing for the AI boom, and SK hynix sells more of it than anyone.

The company is not shy about where the IPO money is going. It plans to pour the proceeds into expanding AI-memory manufacturing. Confirmed projects include the first phase of a fab at the enormous Yongin semiconductor cluster in South Korea, a new advanced-packaging line called P&T7 in Cheongju, and EUV lithography machines (the extremely precise, expensive tools used to etch the tiniest chip features) due by the end of next year.

A US factory too

There is an American piece as well. SK hynix is building its first US production site, a $4 billion advanced-packaging plant in West Lafayette, Indiana, targeted for completion around 2028. The facility is eligible for up to $458 million in CHIPS Act grants and up to $570 million in federal loans. Advanced packaging is the step where individual chips are assembled and connected into a finished product, and it is increasingly where a lot of the engineering happens.

Why the market is so excited

The numbers behind the enthusiasm are striking. SK hynix is reportedly on track to post over 200 trillion won, roughly $133 billion, in operating profit this year. If that holds, employees are said to be in line for bonuses of around $400,000 each. The Seoul-listed stock is up about 220% so far this year and has climbed more than sixfold over the past year.

In late June, SK hynix briefly overtook Samsung as South Korea's most valuable company, closing at around 2,080 trillion won, about $1.35 trillion. That is a notable turnaround for a company that came close to bankruptcy in 2001 and recorded a 7.73 trillion won operating loss as recently as 2023.

A quick caveat worth keeping in mind: the $133 billion profit figure is a projection, not an audited result, and the bonus estimates are reported rather than confirmed by the company. The broader trend, though, is clear enough. Demand for AI memory is running far ahead of supply.

What's next

Here is the detail that says the most about the moment. SK hynix has said its entire 2026 output of HBM, DRAM, and NAND is already sold out, with the crunch expected to stretch into 2027. When a manufacturer has sold next year's production before it has even been made, the constraint is not customers. It is capacity.

That is why $26.5 billion is heading straight into fabs, packaging lines, and lithography gear. The bet is straightforward: AI models keep getting bigger, bigger models need more fast memory, and whoever can build that memory fastest wins. The interesting question for the next couple of years is whether SK hynix can add capacity quickly enough to ease the crunch, or whether rivals like Samsung and Micron close the gap while it tries. Either way, the memory that feeds AI has become one of the most valuable commodities in tech.

If you have ever tried to run an open AI model on your own laptop, you may know the pain: cryptic setup steps, hardware quirks, and documentation written for researchers rather than working programmers. Ollama exists to make that pain go away. Type a command, wait a few minutes, and a model is running. That simple promise has now attracted serious money.

The company just raised a $65 million Series B led by Theory Ventures, on top of an earlier $15 million Series A led by Benchmark's Peter Fenton. That brings total funding to $88 million, not bad for a team of only 14 people.

What Ollama actually does

Ollama, which launched in 2023, helps developers run open-weight AI models directly on their own machines. Open-weight means the model's trained parameters are published, so anyone can download and run them, unlike closed models where you only get access through an API. Ollama also lets you browse and find models, and reach bigger, heavier ones it hosts on its own cloud through subscription tiers from free to $100 a month. Usage there is tracked by GPU time rather than token limits.

The founders, Jeff Morgan and Michael Chiang, previously helped build Docker Desktop, the tool that made shipping cloud apps easier by hiding the hardware headaches. Ollama is essentially the same trick applied to AI. As Morgan puts it, open models arrived in 2023 but were "really hard to use" because they were built for researchers, not builders.

Why it matters now

The numbers are striking. Morgan says Ollama is used by more than 8.9 million developers every month and sits inside 85% of the Fortune 500. On GitHub it has racked up 176,000 stars and nearly 17,000 forks, the platform's rough measures of popularity and how often people copy a project to tinker with it.

Morgan points to a turning point around January, when larger open models became capable of "agentic" tasks like coding, meaning they can carry out multi-step work rather than just answering questions. That shift feeds a growing industry bet: cost-conscious enterprises and startups will lean on cheaper open models for daily work and save pricier closed models for when they really need them.

Fenton, who joined the board, thinks the open-versus-closed debate is framed wrong. "It's not an either/or," he says. But he argues that any company facing high inference costs, the expense of actually running models, has a strong pull toward open-weight options. That trend, conveniently, is good news for Ollama's paid cloud service.

The grumbling in the background

Not everyone is thrilled. About a year ago, some blog and social media posts complained that Ollama's cloud business was pulling focus from the beloved free tool, citing it as an example of the "enshittification" of developer tools, the term for products that get worse as they chase revenue.

Morgan frames the cloud service differently. The best open models are often too big to run on a personal computer, so the company decided to help users find the compute for them. Fenton is blunter: "Nothing has changed for the core product that's free on the desktop."

Worth noting: both Morgan and Fenton declined to share revenue figures or the new valuation, so the business case here rests on user counts and their own read of the market.

What's next

Ollama is part of a wider pattern. AI is spawning a fresh crop of open source projects that turn into venture-backed companies, from inference providers to teams building open models from scratch. Whether Ollama can keep its free fans happy while growing a paid business is the tension to watch. If it manages both, it will have pulled off the Docker playbook a second time. If not, the enshittification critics will feel vindicated.

When people talk about the AI boom, they talk about GPUs, the graphics-derived chips that do the heavy math. The humble CPU, the general-purpose brain of a computer, tends to get ignored. Nvidia would like to change that. It is now pitching its upcoming Arm-based Vera CPU with a phrase it seems rather proud of: a "max single-threaded CPU at scale."

That is a mouthful, so let's translate. A single thread is one continuous line of work a processor handles start to finish. Most big server chips brag about having lots of cores, meaning they can juggle many tasks at once. Nvidia is arguing that for AI, raw juggling isn't the point. Doing one thing fast is.

Why speed beats sheer numbers

Here is the logic. Modern reasoning AI works step by step. It runs the model, looks at the answer, then runs it again, over and over until it lands on a result. Each step depends on the one before it. You cannot start step two until step one is done, so throwing more cores at the problem doesn't help. What matters is how fast a single thread can chew through each step.

Agentic AI, where software "agents" carry out multi-step tasks on your behalf, has the same bottleneck. Agent B cannot begin until it knows what agent A did. It's a relay race, not a crowd. The speed of the runner matters more than the size of the crowd.

Vera is Nvidia's answer. It packs 88 cores (176 threads with a technique that lets each core handle two tasks) on a single monolithic chip, meaning one continuous piece of silicon rather than several smaller ones stitched together. Nvidia says that design keeps things predictable and well fed with data, quoting 1.2 TB/s of memory bandwidth and 3.4 TB/s of core-to-core bandwidth, the latter of which it claims is triple any other data center CPU.

The chiplet tax

That monolithic choice is a deliberate jab at rivals. To reach very high core counts, chips like Intel's Xeon and AMD's Epyc often use chiplets, smaller chip pieces linked together. Nvidia calls the downside a "chiplet tax": uneven memory access and inconsistent performance. Its single-die design, it says, sidesteps that. There is also a physics catch it points out. The more cores you cram in, the slower each one has to run to keep heat and power in check. Great for bulk work, less great for tasks you need finished right now.

Nvidia backs the pitch with customer numbers. Perplexity reportedly saw a 1.5x speedup in coding agent work and 1.9x running concurrent sandboxes. Database firm Starburst claims a 3x uplift in large-scale SQL analytics, and Redpanda reports a 6x drop in latency. Broadly, Nvidia says Vera is 1.8x faster than x86 competitors on agentic workloads.

Take the numbers with salt

Now the honest part. These are vendor-supplied benchmarks, and vendor benchmarks always deserve skepticism. Nvidia doesn't say exactly which Intel Xeon and AMD Epyc chips it compared against, though mid- to high-end models are a safe guess. And AMD has already pushed back on earlier independent tests by Phoronix, claiming a 3.3x advantage of its own when measured across a full 100 kW rack of hardware. Different framings, wildly different conclusions. That is normal in a field where a single day can move trillions of dollars.

What's next

Nvidia isn't standing still. It has already revealed a next-generation Arm core called Rigel, shipping in a future chip named Rosa, which it says will push per-core performance higher than Vera through better instruction handling, more cache, and improved memory management.

The bigger idea is worth watching. If agentic AI really does hinge on single-thread speed the way responsive personal computers always have, then Nvidia's design philosophy makes sense rather than being marketing dressing. The open question is whether Intel and AMD agree. If they start pitching "max single-threaded CPUs at scale" of their own, we'll know Nvidia framed the debate correctly.

AI chips have a feeding problem. The processors that run large models have gotten so fast that memory can't shovel data at them quickly enough, a chokepoint engineers politely call the "memory wall." That is why the current darling of AI hardware, high-bandwidth memory (HBM), is both essential and infuriatingly expensive to build.

A newly published Intel patent application, spotted by the account Underfox, hints at how Intel might sidestep some of that cost. It describes an architecture called cross-batch memory, or XBM, and it takes aim squarely at the parts of HBM that make it pricey and hard to scale.

What HBM does today

Standard HBM stacks memory chips vertically, threads them together with tiny vertical wires called through-silicon vias, and talks to the processor across a silicon interposer. That interposer is a slab of silicon that acts as a fancy circuit board, carrying an extremely wide connection of around 1,024 wires per stack. All that width is how HBM moves so much data. It is also what makes packaging it a costly, finicky affair, since every one of those wires has to be routed just so.

What Intel is changing

XBM makes two big moves. First, it changes where the memory lives. Normal DRAM cells sit in the base silicon layer where transistors are usually built. Intel instead builds the memory cell higher up, in the metal-and-wiring layers above the transistors, using thin-film transistors. That approach, called back-end-of-line fabrication, lets Intel chop each die into many small, independently addressable blocks. The filing describes dies of roughly 1.5 gigabytes each, stacked eight high and scaling to sixteen.

Second, and more consequentially, it drops HBM's wide parallel connection. Instead of a thousand-plus wires crossing an interposer, XBM funnels data out over UCIe links running at 32 gigatransfers per second. UCIe (Universal Chiplet Interconnect Express) is an industry-standard way for chip building blocks to talk to each other, and using it makes the design "chiplet-native." Intel's argument is that a standard serial interconnect is simpler and cheaper to package than an interposer-bound stack.

There is a catch worth flagging: 32 gigatransfers per second is UCIe's current top speed. In other words, XBM would launch already pressed against the spec ceiling, with no obvious room to go faster.

Fixing defects after the fact

Tall memory stacks are hard to build without flaws, and a single bad cell can sink an otherwise good chip. Intel leans hard on repairability here. The base die at the bottom of the stack carries spare channels, built-in self-repair logic, and redundant memory arrays that can stand in for defects in the dies above. The idea is to claw back manufacturing yield after assembly rather than throwing away expensive stacks.

A surprising chunk of the filing is not about the memory at all, but about how to mount it. Intel details packaging tricks to shrink the stack's height and remove a stiffening component normally needed to stop the package from warping. That is the concrete basis for its "smaller, cheaper package" pitch.

One of two bets

XBM should not be confused with ZAM, a separate memory architecture Intel is co-developing with SoftBank subsidiary SAIMEMORY and plans to present at the VLSI Symposium 2026, with commercialization aimed at 2029. ZAM keeps largely conventional DRAM but innovates on how the layers are bonded together. Read side by side, the two suggest Intel is hedging with at least two HBM alternatives. Fitting, perhaps, for a company that started life as a memory maker back in 1968.

What's next

Temper the excitement. This is a patent application filed in December 2024, not a product or even a roadmap. It signals intent, not a shipping part. Backend-transistor DRAM remains unproven at manufacturing scale, the UCIe interface is already maxed out, and XBM would still have to prove itself against newer HBM generations and Intel's own ZAM timeline. Patents are cheap; competitive memory is not. Still, the fact that so many players are attacking the interface rather than the logic tells you where the real bottleneck now sits.

Teaching a robot to pick up a red cube sounds simple until you realize the robot has no idea whether it succeeded, no sense of what happens next, and no easy way to learn from its own mistakes. Hugging Face's LeRobot project, an open toolkit for robot learning, just shipped version 0.6.0 with a clear theme: close that loop. The update focuses on robots that imagine, evaluate, and improve.

Policies that imagine the future

The headline additions are three "world model" policies. A world model is software that learns to predict what a scene will look like a moment from now, so the robot can, in effect, rehearse before it moves.

The clever part is cost control. VLA-JEPA trains a small model to anticipate upcoming frames, then throws that machinery away at inference time, so you get the benefit of imagination without paying for it when the robot actually runs. LingBot-VA predicts future video and actions together and can even save what it imagined so you can compare it with reality; its inference fits on a single 24 to 32 GB GPU. FastWAM pairs a video-generation expert with a compact action expert, learns to "dream" its own rollouts during training, and then skips the dreaming entirely when it acts.

A growing zoo of action models

LeRobot also expanded its lineup of VLAs, short for vision-language-action models, which turn what a camera sees plus a plain-English instruction into robot movement. NVIDIA's GR00T integration jumps to the newest N1.7 generation, tested against NVIDIA's original code to produce identical outputs. The Allen Institute for AI's MolmoAct2 arrives ready to run zero-shot on hobbyist SO-100 arms in about 12 GB of memory. And EVO1 makes the case that VLAs need not be enormous, packing everything into 0.77 billion parameters that run in real time on modest hardware.

Knowing when the robot actually succeeds

Here is a gap most people never think about: robots are bad at grading their own homework. v0.6.0 adds a unified reward models API, where a reward model watches a task and scores progress and success. Robometer, built on Qwen3-VL-4B and trained on more than a million robot trajectories, scores any dataset from raw video and a language instruction with no task-specific tuning. TOPReward goes further and uses no special weights at all. It simply asks an off-the-shelf language model how likely the word "True" is given the video and the task. Any capable model becomes a success detector.

Turning failures into fuel

The new lerobot-rollout command handles deployment and supports DAgger-style corrections, meaning a human can step in when the robot fumbles and those corrections flow straight back into the training data. Combined with the reward models, this is the actual learning loop: act, judge, fix, repeat.

Supporting all this is a batch of practical upgrades. Datasets can now record depth from an Intel RealSense camera, store rich timestamped language annotations generated automatically by a language model, and use whatever video codec you prefer. Video training loads up to twice as fast, and pulling a subset of a large dataset dropped from around 275 seconds to a fraction of a second in Hugging Face's own benchmark. You can also train models bigger than a single GPU using a technique called FSDP, or offload the whole job to the cloud.

Measuring it all

To check whether any of this actually helps, v0.6.0 unifies six new simulation benchmarks under one command. They range from LIBERO-plus, which throws roughly 10,000 perturbed scenarios at a policy to find where it breaks, to RoboCasa365's 365 kitchen tasks across 2,500 generated kitchens, to memory and long-horizon reasoning exams. Each ships with a baseline checkpoint that is smoke-tested automatically.

What's next

The open question the release keeps returning to is whether imagining the future genuinely makes robots better, or just more expensive to train. By putting world models, reward models, and shared benchmarks in one place, LeRobot is less interested in declaring a winner than in making the experiment easy to run. That is the quieter kind of progress worth watching: not a single breakthrough, but the plumbing that lets everyone test their own.

It was a week about the invisible plumbing behind AI: the rules that govern it, the power that feeds it, and the factories that build the chips. Three very different stories, one shared thread. Let's dig in.

First, a disappearing act with a tidy ending. Anthropic Restores Claude Fable 5 tells the story of a capable AI model that vanished worldwide for 18 days after a US export ban. Since there was no easy way to check who was typing into the chat box, Anthropic pulled it everywhere, then brought it back on July 1st with a single safety filter tuned to block one prompt.

Next, a genuinely wild live demo. In Startup Powers an Nvidia PC With Nuclear, a team from Valar Atomics plugged a Blackwell desktop into a microreactor on stage and switched it on. The PC served a live website that, they say, stays online only as long as the reactor keeps running, a hint at how AI factories might sidestep strained grids.

Finally, a quieter but meaningful win. Report Claims Intel Fixed 18A Yield Issue unpacks a research note suggesting Intel cleared a recurring problem on its newest chip process. It's good news for the comeback story, though consistent still isn't the same as good enough.

Next week we'll keep watching where the rules, the power, and the silicon meet. See you then.

Making advanced chips is less like printing money and more like baking bread at industrial scale. Some batches come out fine, others don't, and figuring out why is half the job. Intel appears to have solved one of those recurring problems on its newest manufacturing process, according to a research note that could matter a lot for its comeback story.

The claim comes from BlueFin Research Partners, an investment research firm, and was shared via social media rather than announced by Intel. So treat it as an unofficial signal, not confirmed fact. The note says Intel has resolved a "wafer-to-wafer yield issue" on its 18A process and is ramping production to 12,000 to 15,000 wafers per month at two sites.

What 18A and the fix actually mean

A quick vocabulary check. A wafer is the round silicon disc that gets carved into individual chips, called dies. "18A" is Intel's name for its latest, most advanced production process, roughly in the 1.8-nanometer class. "Yield" is the share of usable chips you get from a wafer. Higher yield means more sellable chips and better economics.

Wafer-to-wafer variability is exactly what it sounds like. Good wafers and poor wafers come off the same line, making output unpredictable. Fixing it means production is now more consistent from wafer to wafer and lot to lot.

Here is the important caveat. Consistency is not the same as high quality. Yield depends on several things: defect density, differences between the center and edge of a single wafer, whether chips hit their speed and power targets, and packaging. Resolving wafer-to-wafer variability tackles just one of those. Intel has previously said 18A yields improve about 7% per month, so steadier production mainly means it can now march toward its targets on a predictable schedule, not that it has already arrived.

Capacity is building

The report also says Intel now has capacity for roughly 30,000 wafer starts per month, split across its D1X development fab in Oregon and its Fab 52 high-volume plant in Arizona. That is a solid figure this early in a ramp.

There is a wrinkle worth flagging. Running high-volume manufacturing inside a development facility, which D1X is, costs more than using a fab purpose-built for mass production. And without numbers on overall yields, it is hard to say whether Intel can churn out enough of its upcoming chips, including Core Ultra 3 "Panther Lake" laptop processors and "Clearwater Forest" Xeon server chips.

What's next

Intel plans to repeat this approach with its next process, called 14A, in the 1.4-nanometer class. According to BlueFin, D1X in Oregon will again serve as the first high-volume site, with the first phase of Intel's new Ohio plant becoming the second. Intel has confirmed it aims to start high-volume 14A production in 2029, while the Ohio facility's first phase is expected to complete in 2030 and come online between 2030 and 2031.

The bigger picture is simple. Intel's future as both a chip designer and a contract manufacturer hinges on getting these advanced processes to work reliably and affordably. A steadier 18A line is a meaningful step, assuming the report holds up. The real test will be whether defect rates, performance targets, and costs all line up well enough to make the chips profitable. Watch for official yield figures and for how quickly Panther Lake and Clearwater Forest actually reach buyers.

At a live event, a team member from startup Valar Atomics plugged an Nvidia RTX desktop PC into a nuclear reactor, turned the reactor up to 37% of its power, and switched the machine on. That PC then served a live website, nuclearwebsite.com, which the company says stays online only as long as the reactor keeps running.

What it is

Valar Atomics activated its Ward 250 microreactor and used it to power a single Nvidia Blackwell chip. Blackwell is Nvidia's current generation of AI processor. A microreactor is a very small nuclear reactor, and a small modular reactor (SMR) is a compact, factory-built design meant to be deployed faster than a full-scale plant.

CEO Isiah Taylor walked the audience through the chain of events. Uranium atoms fissioning in the reactor hall produce about 100 kilowatts of thermal energy. A pressurized helium cooling loop carries that heat to a thermal electric generator, which converts it into electrical current. That current powered the Nvidia chip on stage.

The company also announced a partnership with Nvidia to build a 30-megawatt, closed-loop AI factory. The "closed loop" detail matters: the design is meant to avoid drawing on local water supplies.

Why it matters

Data centers have become a live political issue. They are being blamed for sharp rises in power and utility bills, higher water consumption, and a drop in quality of life for nearby communities. According to the source, 7 out of 10 Americans say they do not want a data center in their backyard, and that resistance helped delay or cancel at least 75 projects in the first quarter of 2026 alone.

That pushback is pushing companies to find power that does not lean on the existing grid. Reactors like Ward 250 could, in principle, supply electricity directly to AI sites without adding strain to national or local networks. Water is the other half of the pitch. Amazon, Microsoft, and Nvidia are also working on approaches meant to cut data center water use by as much as 100%.

Big AI players have been circling nuclear for a while. Amazon, Google, Microsoft, Nvidia, and Oracle all began investing in nuclear technologies as early as 2024, betting that AI data centers would need enormous amounts of power.

Read the claims carefully

Valar Atomics says it is the first startup to achieve power production. That claim deserves context. The Department of Energy notes that two other firms, Deployable Energy with its Unity reactor and Antares Nuclear with its Mark-0, have also achieved criticality. Criticality means a reactor has reached a self-sustaining chain reaction, an early step toward generating electricity. In other words, several startups are moving in the same direction, and the exact order of milestones is contested.

Scale is worth keeping in perspective too. The on-stage demonstration involved 100 kilowatts of thermal energy powering one chip. The proposed AI factory is 30 megawatts. That is a large jump from a single desktop to a facility, and the source describes the factory as a plan rather than a finished build.

What's next

The immediate test is whether Valar Atomics and Nvidia can move from a stage demonstration to a working 30-megawatt facility that actually runs without tapping local water. If closed-loop, reactor-powered AI sites prove practical, they could reshape where and how data centers get built, especially in communities that have started saying no.

For now, the useful signal is not the spectacle of a PC running on nuclear power. It is that grid strain and water use have become real obstacles to AI expansion, and companies are now treating on-site energy as part of the hardware problem, not an afterthought.

For 18 days, one of Anthropic's most capable AI models simply vanished. On June 12th, the U.S. Department of Commerce placed export controls on Claude Fable 5, barring any foreign national—including Anthropic's own non-citizen employees—from using it. With no reliable way to check the nationality of everyone typing into a chat box, Anthropic pulled the model worldwide. On July 1st, it came back.

What ended the standoff wasn't a sweeping overhaul. It was a single safety filter, narrowly tuned to block one technique. That small fix, and what it reveals about how governments and AI labs now negotiate access to frontier systems, is the part worth watching.

What actually happened

The trouble started when Amazon researchers found a way to prompt Fable 5 into identifying software vulnerabilities—weaknesses in code that attackers can abuse—and, in one case, writing code showing how such a flaw could be exploited. That prompted Commerce to impose export controls on Fable 5 and on Mythos 5, the more capable model it's built on.

Anthropic's response was to train a new classifier—a filter that recognizes a specific type of request—to catch the technique Amazon flagged. The company says it blocks that prompt in more than 99% of cases and reroutes flagged requests to an older model, Opus 4.8. Commerce's own Center for AI Standards and Innovation reviewed the safeguards before lifting the controls. Fable 5 is now returning across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork, with cloud providers AWS, Google Cloud, and Microsoft Foundry to follow.

The filter blocks the prompt, not the ability

Here's the important nuance: the classifier targets the reported request, not the underlying capability. Fable 5 can still identify the vulnerabilities in Amazon's report—the filter simply detects the request and reroutes it rather than removing the skill from the model. As a side effect, the change also catches some harmless coding and debugging requests.

That design has an obvious limit. Detection-based safeguards are exactly what were defeated to trigger the ban in the first place. A filter tuned to one known technique does nothing for techniques nobody has found yet. Anthropic concedes as much, acknowledging that no model can be made fully resistant to jailbreaks—attempts to trick a model into ignoring its safety rules—and that it expects more to surface.

Why the capability may have been oversold

A joint review by Anthropic, the government, and Amazon complicates the case for singling out these models. It found that Opus 4.8, OpenAI's GPT-5.5, and China's Kimi K2.7 could all identify the same vulnerabilities. Every model tested—including Haiku 4.5, Sonnet 4.6, and several Opus versions—could reproduce the single exploit demonstration. In other words, the cyber capabilities that justified the controls appear to be widespread across the field, not unique to Mythos-class systems.

The offline period carried a competitive cost, too. While Fable was gone, Chinese lab Z.ai's GLM-5.2 held top benchmark positions by default, including the leading accessible score on the AA-Briefcase multi-week task test. Fable 5's return reclaims those spots. Mythos 5, which carries fewer guardrails and stays limited to Project Glasswing partners, went back to a set of U.S. organizations on June 26th.

What's next

Anthropic has opened a HackerOne program inviting outside researchers to report new Fable 5 jailbreaks, and it committed to giving designated government partners earlier access to test future frontier models before public release. That last point may be the most consequential: a precedent for pre-release government review of powerful models.

The episode shows how quickly a single flagged prompt can pull a model off the global market—and how a narrow fix can put it back. But the deeper question lingers. If the risky capability is common across today's leading models and no filter fully stops jailbreaks, export controls aimed at one model may prove a blunt tool for a problem that lives across the whole industry.

• About Tempus
• Learn - a couple of courses to further your knowledge in AI
• AI Jobs - a listing of fresh jobs related to AI
• In Other News - a few interesting developments we're tracking

Tempus AI is a health tech company that uses artificial intelligence to help doctors treat patients more precisely—especially those with cancer, heart disease, or mental health conditions.

Think of it as a super-smart medical assistant that can read tons of data (like DNA tests, X-rays, and medical records) and then give doctors helpful advice about what treatments might work best.

Real-Life Examples:

Finding the Right Cancer Treatment

Imagine a woman with breast cancer. Doctors send her tumor for testing to Tempus. Tempus analyzes her DNA and RNA using AI and discovers she has a rare mutation. The platform then suggests a specific drug that works well for that mutation—even if her doctor didn’t know about it.

Catching Heart Problems Early

Tempus worked with Northwestern Medicine to create an AI tool that looks at ECG (heart test) results. The AI can spot people who may soon develop atrial fibrillation—a serious heart rhythm problem—before symptoms appear.

Helping Patients Understand Their Health

In 2025, Tempus launched a digital assistant called Olivia. It helps patients collect all their health info (like blood pressure, imaging scans, medications), and then uses AI to answer questions like:

“Why am I taking this medication?”
“Can I join any clinical trials?”

It’s like having ChatGPT for your medical records.

Why It Matters

For doctors: They get better tools to treat patients faster and more accurately.

For patients: They get personalized care and feel more in control of their health.

For researchers: They can discover new drugs faster by analyzing Tempus’s massive health database.

Where It Is Used

Hospitals like Cleveland Clinic and Northwestern Medicine

Pharma companies like Pfizer and AstraZeneca (to find new drug targets)

Patients in the U.S., and soon, Japan (through a partnership with SoftBank)

Tempus AI turns big, messy health data into clear, useful answers that save lives.

📚 Learn

🧑‍💻 Jobs

🔔 In Other News

For some in the industry, AI filmmaking is already becoming mainstream

“It’s being used by everybody that doesn’t talk about the fact that they’re using it,” Michael Burns, vice chairman of Lionsgate, said at a panel conversation at Runway’s AI Film Festival.

NBC NewsAngela Yang

Sam Altman says Meta tried and failed to poach OpenAI’s talent with $100M offers | TechCrunch

OpenAI CEO Sam Altman said that Meta tried to poach its employees with nine-figure offers, but failed to recruit OpenAI’s best people.

TechCrunchMaxwell Zeff

Samsung, Nvidia are investing $35 million combined in THIS robotics startup. Here’s why | Mint

Samsung and Nvidia are set to invest, $10 million and $25 million respectively, in Skild AI to enhance their presence in consumer robotics space. Part of a $100 million Series B funding round led by SoftBank, this will value Skild at $4.5 billion.

mintJocelyn Fernandes