Google Turns TPU Capacity Into a Blackstone-Backed Neocloud

Google’s AI infrastructure strategy is moving outside the walls of Google Cloud. Ed Ludlow described the new Google-Blackstone venture as a neocloud company built around Google’s homegrown Tensor Processing Units, with Blackstone contributing an initial $5 billion of equity capital, a target of 500 megawatts of computing capacity by 2027, and debt financing that Caroline Hyde said could leverage the commitment up to $25 billion.

The immediate market read was competitive pressure on existing neoclouds. Ludlow pointed to declines in CoreWeave, IREN and Nebius, describing them as Nvidia-focused architectures facing a new rival that would use Google chips instead.

Mandeep Singh said the logic is clearest when compared with Nvidia’s ecosystem. Nvidia has a large roster of neocloud players. Google, by contrast, has comparatively few external TPU providers; Singh cited TerraWulf and Cipher Mining, but said “there aren’t a lot” of TPU providers. If Anthropic wants to use “almost $200 billion of capacity” from Google, he argued, Google needs a way to ramp capacity beyond its own cloud.

The Blackstone structure matters because it shifts more of the data-center burden away from Google. Singh said that in Google Cloud, customers are served from Google’s own data centers. In the new venture, Blackstone is expected to handle power, cooling and land, while the data centers run Google TPUs. That gives Google a path to live capacity sooner than if it were trying to expand only through its own facilities.

Bloomberg Intelligence’s on-screen note framed the pact as a boost to Google’s chips business as well as its cloud business, saying the joint venture showed Google’s focus on ramping up TPUs “in addition to its cloud segment,” where the company has seen faster growth than Microsoft Azure and Amazon AWS.

The distinction is important: this is not just another cloud availability zone. It is a financing and deployment model for getting Google silicon into a market where Nvidia-centered neoclouds have so far defined much of the category.

The labor story around AI was presented less as a single trend than as a contested interpretation. Meta is moving 7,000 workers into AI-related roles ahead of job cuts, according to an internal memo cited by Bloomberg. The new corporate structure will be “flatter” and organized around smaller teams, including groups focused on AI-related products such as agents and apps. The same memo said Meta is cutting 10% of workers, roughly 8,000 people.

Standard Chartered CEO Bill Winters supplied the bluntest formulation. In a quote shown on screen, he said the bank’s changes were “not cost cutting,” but rather “replacing in some cases lower-value human capital with the financial capital and the investment capital we’re putting in.” Hyde called “lower-value human capital” an extraordinary choice of words.

Marta Norton was cautious about treating these announcements as proof of broad AI-driven layoffs across the economy. She said many companies added labor after the pandemic and are now rightsizing. In some cases, she suggested, companies may be attaching the AI label to workforce reductions to make them seem more forward-looking — “to maybe AI whitewash it.”

Norton said she is watching hiring more closely than layoffs as a signal of AI’s macro impact. The more important question, in her view, is whether firms become slower to hire than they otherwise would because AI tools can absorb some work.

On adoption, Norton described the first phase as widespread but still limited. Employees are using chatbots for “low-value research,” she said, but the more consequential productivity gains would come from agent deployment. Companies are beginning to structure workflows around agents, though she emphasized that this remains early.

Nvidia’s earnings were framed as the market test for whether AI demand is translating into enough profit to justify the rally. Jensen Huang told Ludlow in a clip from Las Vegas that Nvidia’s supply chain is lined up, including silicon photonics, but “the demand is much greater than the overall capacity of the world.”

Norton said that in a supply-constrained environment, investors still need evidence of monetization: demand should be moving earnings. She also pointed to constraints that could affect Nvidia itself, including memory chips, and said gross margins and supply navigation would matter. But she agreed with Huang’s broader view that the AI build remains in “inning one, inning two,” with a supply-demand mismatch likely to last several quarters or years.

The jury in Elon Musk’s case against his OpenAI co-founders did not decide whether OpenAI’s restructuring violated the commitments Musk alleged. It decided that Musk waited too long to sue.

Madlin Mekelburg called the verdict “a bit anticlimactic” because the jury did not reach any of Musk’s substantive claims. Instead, it found that he had not proven his case within the statute of limitations for the specific claims he brought. In effect, the jury concluded that Musk had concerns about the conduct and should have known about it years before he filed suit.

A post from Musk shown on screen said: “Regarding the OpenAI case, the judge & jury never actually ruled on the merits of the case, just on a calendar technicality.” Mekelburg said it was unsurprising that Musk planned to appeal, but that an appeal on statute-of-limitations grounds would be narrower than an appeal after a merits ruling.

Dorothy Lund explained the legal structure. The two claims, she said, were that OpenAI, Greg Brockman and Sam Altman violated a promise to Musk that OpenAI would maintain a permanent charitable mission to develop safe, open-source AI technology, and that they received undeserved benefits because of those broken promises. OpenAI’s response was that Musk filed too late: he brought the case in August 2024, while the statute of limitations was three years, and the conduct he complained about occurred in 2018 and 2019.

Musk’s argument, Lund said, relied on tolling: if harm has been concealed or not discovered, the limitations period can be paused. Musk argued that Altman and OpenAI had concealed what they were doing and that he did not truly understand the harm until 2023. That is why the issue went to the jury. Jurors had to assess testimony and documents to determine when Musk knew, or reasonably should have known, about the harm he alleged.

The unresolved issue is the larger one that drew public attention: whether OpenAI violated its charitable purpose when it created a for-profit affiliate. Lund said the appeal will not resolve that either. Because the limitations issue turned on factual findings, she said it will be difficult to overturn. Appellate courts rarely reverse a jury’s fact-intensive findings, especially where the judge also reviewed the evidence and reached the same conclusion.

For Musk to move forward, Lund said, the likely path is not relitigating the merits but arguing that legal or procedural errors prejudiced the outcome — for example, improper jury instructions or wrongly decided evidentiary rulings.

Parag Agrawal is trying to replace flat-fee AI content deals with a marketplace that pays publishers and data providers according to how much their content helps an AI agent complete a task.

Agrawal said Parallel has been building toward this model for two and a half years, starting from the premise that agents will use web content “thousand X more than humans ever have.” If agents become the primary users of the web, he argued, the existing business models — advertising and subscriptions — no longer work. The challenge is to align agents doing real work with the content owners whose material feeds those agents.

Parallel’s new product, Index, is meant to do that. Ludlow framed the problem as a missing marketplace: agents need web data and proprietary datasets, but the system for pricing that data is underdeveloped. Agrawal contrasted Index with current deals between large AI labs or hyperscalers and large publishers, which he described as flat-fee arrangements. Over time, he said, those deals do not let content owners participate in the growing economy of agents.

Agrawal said agents are already doing real knowledge work for customers building AI scientists, AI lawyers and finance tools. If those agents produce value, he argued, the data sources that made the work possible should participate in that value. Index is designed so that higher-quality data gets paid more, data used in higher-value work gets paid more, and content owners grow as agents perform more valuable tasks.

The mechanism he emphasized is Shapley value, a game-theoretic concept for dividing value among contributors to a cooperative outcome where the whole is greater than the sum of the parts. Agrawal said he had been “obsessed” with Shapley values for three years because they offer a way to assign marginal contribution: if an AI agent completes valuable work using someone’s data, the system can estimate how much that data contributed.

The Index website shown on screen framed the problem starkly: “The web’s primary user is changing. AI agents are displacing the attention that built the content economy, without replacing the revenue that sustained it.” A demo showed domain-level scores for impressions, citations, value and uniqueness, including nyc.gov, nih.gov, loc.gov, nasa.gov and sf.gov. It also showed citation share and unique value by domain, with nih.gov capturing the largest citation share among the displayed domains.

Parallel’s launch partners include premium publishers and data providers. Agrawal named The Atlantic, Fortune, PR Newswire, PitchBook, Tracxn and ZoomInfo, as well as independent creators including Alex Heath and Azeem Azhar. The displayed partner graphic also included Enigma and Fiscal.ai. Agrawal said the goal is for content owners large and small to participate, and for agents built by companies other than the largest providers to access the content.

When asked about X, formerly Twitter, Agrawal said he still uses it and still calls it Twitter. He described the platform built during his years there as enduring in value because it supported open content access rather than permissioned groups. He tied that directly to Parallel’s mission: broad content access for everyone’s agents, not just a few.

Meta’s planned AI facility in Richland Parish, Louisiana, was described as an economic shock arriving in a struggling rural region. Riley Griffin said the headline $200 billion figure is largely about the chips inside the facilities, not necessarily direct spending in the community.

The planned campus spans nearly 4,000 acres. Griffin said the facility would involve 5 gigawatts of compute capacity and another 2.5 gigawatts to support the broader campus. President Donald Trump, after speaking with Mark Zuckerberg, described it as a Manhattan-sized project.

Griffin’s reporting focused on what the project means for Richland Parish and the Louisiana Delta, a region she said has struggled for decades. She noted that farmers there are losing $400 on every acre of cotton they sow. Caroline Hyde cited Dustin Morris, a soybean and corn farmer, as an example of the local dilemma: flying over his own land and trying to decide whether the moment has come to sell.

Griffin described an expected land rush. As very large data centers shift into rural areas, particularly in the South, landowners are asking whether to trade crop land and family history for data-center development. She said residents see Meta as a possible lifeline, but the question is what kind of lifeline it will be.

Meta negotiated the deal privately over the course of a year, Griffin said. The company says it wants to be engaged with the community and is investing in schools. There is a school at the corner of the data-center site, and Griffin said she had heard Meta was reviewing plans amid discussions over whether it should be moved.

The central uncertainty is jobs. Griffin said Meta is bringing only 500 jobs to a community that had sought 5,000 from an auto manufacturer over the years and failed to attract that investment. Her question was whether 500 jobs can provide a long-term economic lifeline.

At JPMorgan, AI adoption is speeding up not only because models are improving but because the bank is finding more ways to apply them. Lori Beer, JPMorgan’s global CIO, said the last six months have rapidly accelerated both model creation and practical application. The question, she said, is how to drive and adapt to change quickly.

Beer rejected a simple jobs narrative. She said JPMorgan has seen productivity opportunities, including 10%, 20% and 30% improvements from the first generation of AI tools in technology work. Agentic systems should produce more. But she also emphasized demand: the bank needs to create new products, services and customer experiences, and AI is also raising cybersecurity risk, which requires additional investment.

Cybersecurity was one of Beer’s clearest examples of AI’s dual role. Asked about Mythos, she said AI can help understand vulnerabilities and, if integrated safely, improve software development. But the same models also increase risks. The task is to capture and clear vulnerabilities, update systems securely, and use models to protect customers, identify fraud and spot issues faster.

Beer said the technology is being embedded horizontally across the stack, but the harder problem is change management and leadership. Her management team spends weekends building their own apps with coding-agent tools so they can understand what is coming. Leaders, not only software engineers, need to understand the shift, she said, because organizations must pivot under uncertainty without losing sight of customer and client needs.

The build-versus-buy calculus is also changing. JPMorgan has tried to concentrate its engineers on things that are competitively differentiating and valuable to customers. As software engineering becomes faster and more cost-effective across the product-development life cycle, Beer said the trade-offs will evolve. But the bank will still buy enterprise software for functions that help run the bank without differentiating its customer products.

On banking’s longer-term shape, Beer said the sector is already transforming. The obligation that does not change, in her telling, is trust: JPMorgan must protect customers and clients and serve their needs across more than 100 countries. The products, services and delivery mechanisms will evolve.

Armada’s pitch is that not all AI infrastructure belongs in centralized hyperscale data centers. The company raised $230 million in an oversubscribed Series B round at a $2 billion valuation and announced a partnership with Johnson Controls to produce modular data centers.

Dan Wright described Armada’s modular data center as a “full stack AI factory” that can be deployed anywhere. The company’s framing is “speed, scale, and sovereignty”: deploy within weeks, scale with demand, and keep data sovereignty down to the site level. Wright said data sovereignty used to be discussed in terms of countries, but increasingly sophisticated cyberattacks make site-level sovereignty important.

The round was co-led by Overmatch, 8090 Industries and BlackRock, according to Wright. He said the money will help accelerate deployment of the “US AI stack,” and that the Johnson Controls agreement will begin with Galleon Forge 1 in Arizona for continuous manufacturing of modular data centers. Wright compared the modules to shipping containers: transportable, deployable globally and compatible with any source of power.

That last point is central to Armada’s thesis. “Energy is distributed,” Wright said. “Compute should also be distributed.”

Demand, he said, is coming from energy companies and defense. Energy companies maintain a hard separation between IT and OT, meaning AI infrastructure often needs to run on rigs and at refineries, behind the firewall and at the edge. Wright said those companies are trying to become fully autonomous over the next few years, which requires local compute.

Defense is another driver. Wright said that during the conflict in the Middle East, an ally called needing one of Armada’s systems immediately; Armada deployed within days and shipped more when asked.

Armada’s partner ecosystem includes Microsoft, OpenAI, Nvidia, Dell and Palantir. Wright said a recent Microsoft partnership brings Foundry and Azure Local to customers that want to run Microsoft services and models at the edge, including air-gapped environments. He described Armada as “the hyperscaler for the edge,” doing for edge infrastructure what hyperscalers did for centralized cloud: assembling services and making them accessible to more customers.

Asked whether the broader move toward on-prem and edge AI is an addressable market, Wright said yes. He expects “sovereign AI factories,” meaning every country and every company will have its own AI factory. To do that, he said, customers need chips, servers and infrastructure; Armada is trying to supply the infrastructure layer.

Several shorter items reinforced the same market backdrop: capital, hardware, talent and regulatory positioning remain in motion around AI.

Moonshot AI, developer of the Kimi chatbot, is restructuring to comply with Beijing’s rules and prepare for a potential Hong Kong IPO. Chinese robotics company Linkerbot is also exploring a Hong Kong IPO that could come as early as this year, targeting a $6 billion valuation. Hyde said Linkerbot makes dexterous robotic hands and counts Samsung and Tencent among its clients. A Barclays report shown on screen said humanoid robots could offset as much as 60% of China’s labor decline by 2035.

On the U.S.-China technology relationship, Huang said H200 chips are licensed to sell to China, but the Chinese government must decide how much of its local market to protect and how much to expand with more AI capacity. He said demand in China is “incredible” and that agentic AI is making progress there as well. He did not discuss H200 sales directly with Chinese officials, he said, but traveled to represent and support President Trump. Michael Dell said he hoped for more economic collaboration between the United States and China, arguing that it would lead to better outcomes and prosperity.

Apple’s hardware organization is also changing. Ludlow reported that Apple chief hardware officer Johny Srouji is reorganizing hardware development and shifting oversight of key functions, including product design, in an effort to speed future-device work and better integrate in-house silicon teams with product teams. Shelly Goldberg and Dave Pakula were named as new overseers of product design.

The day’s AI talent news came from Andrej Karpathy, a founding OpenAI member, who said in a post on X that he had joined Anthropic. The post shown on screen said he would work on R&D and help train new AI models, while remaining passionate about education and planning to resume that work in time. Ludlow called the move “a bit of a wow” in the AI talent market.