MARK ZUCKERBERG WANTS YOUR CLOUD DOLLARS: META PLANS TO SELL AI COMPUTE POWER AND CHALLENGE AMAZON, GOOGLE, AND MICROSOFT ON THEIR OWN TURF
Mark Zuckerberg has been spending money on AI like a sailor on shore leave who just hit the casino. Not a metaphor. The man committed $182.9 billion to AI infrastructure. For comparison, that number is larger than the GDP of Hungary. For a very long time, the only thing he had to show for it was Meta AI, a chatbot that most people use once and forget, and Llama, an open-source model that the whole world uses but nobody actually pays for.
So when Bloomberg reported this week that Meta is building a cloud computing business to sell that excess infrastructure to outside companies, you should not have been surprised. What is surprising is that it took this long.
Here is how the math works. Meta has built enormous data centers. One in Ohio that Zuckerberg described as the size of Manhattan. More coming in Louisiana. All of it stuffed with Nvidia GPUs that cost tens of thousands of dollars apiece. Those chips need to be doing something around the clock to justify the spend. If your own AI models cannot generate enough demand to keep the lights on, you rent the extra capacity to someone who can use it.
SpaceX figured this out first. In May, SpaceX signed a deal to sell the excess compute at Colossus 1 to Anthropic for $1.25 billion a month. Google signed for $920 million a month. Reflection AI kicked off a $6.3 billion deal on July 1. That is SpaceX turning a data center into a money printer while running rockets on the side. Meta watched this happen and thought: we have the same thing, let us do the same thing.
The new business, reportedly called Meta Compute, will be led by Daniel Gross from Meta Superintelligence Labs and Dina Powell McCormick, the company’s president. They are debating whether to sell raw GPU access like CoreWeave, or hosted AI models like AWS, or both. Muse Spark, Meta’s closed-weight model launched in April, could be offered through this infrastructure, which would mean companies paying Meta to run their AI rather than sending data to OpenAI or Anthropic.
The market loved it. Meta stock jumped nearly 9 percent on the news. That kind of reaction tells you exactly how desperate investors are for signs that hundreds of billions in AI infrastructure spending will eventually produce actual revenue.
Here is the uncomfortable truth the market is pricing in. The AI race has quietly shifted. The winners are not necessarily the companies with the smartest models. They are the companies that own the land underneath the models. The data centers, the chips, the power contracts. Google has it. Microsoft has it. Amazon built its whole empire on it. And now Zuckerberg wants a seat at that table too. After spending $182.9 billion, can you blame him?
Source: TechCrunch | CNBC
ANTHROPIC DROPS CLAUDE SONNET 5 AND DARES RIVALS TO MATCH THE PRICE: THE AGENTIC AI WARS ARE NOW A RACE TO THE BOTTOM
There is a war happening in AI right now, and the primary casualty is the price of intelligence. Anthropic pulled the trigger on Monday when it launched Claude Sonnet 5, a midsize model it claims can do things that, just a few months ago, required the most expensive frontier models.
The pitch is simple enough. Sonnet 5 does agentic work. It browses the web, writes code, uses terminals, makes plans, and runs tasks end to end without needing someone to babysit it every five minutes. It checks its own output without being told to. On the agentic coding benchmark, it scored 63.2 percent, which puts it close to Opus 4.8 at 69.2 percent. Opus 4.8 is the flagship, the thing companies pay serious money for. Sonnet 5 gets you most of the way there for much less.
And Anthropic made it the default model for every free and Pro Claude user. Not an upgrade you unlock by paying more. The new default. Free users who just wanted a decent chatbot are now running an agentic AI powerhouse whether they asked for one or not.
The pricing is the real weapon. Two dollars per million input tokens and ten dollars per million output tokens through August 31, then it steps up to three and fifteen. Compare that to OpenAI’s GPT-5.5, which costs more. Google’s Gemini 3.1 Pro, which costs more. Gemini 3.5 Flash is still cheaper, but Anthropic is staking out territory it has not traditionally owned, which is the affordable end of the market.
This matters because the AI industry is shifting hard from chatbots to agents. Agents are the product now. Not the chat window. And agents run constantly. They hit APIs, loop through tasks, make thousands of small decisions across the day. At enterprise scale, the cost per million tokens is the cost of running your operation. Companies doing that math are looking very carefully at whether they really need to be on the expensive models for every single task.
Zapier tested it on a two-part workflow and said it handled an update to Salesforce account tiers and a launch announcement email end to end without stopping partway. Their senior engineer called it “a no-brainer for day-to-day automation.” That one line is more useful to Anthropic’s enterprise sales team than any benchmark chart.
The safety numbers are also better. Anthropic says Sonnet 5 refuses bad requests more cleanly, hallucinates less, and is harder to push around with prompt injection attacks than its predecessor. In an agentic world where these models are touching live systems, that is genuinely important.
The timing of all this is not accidental. Anthropic has filed confidentially to go public. Making Sonnet 5 the free default is a land-grab for users before the IPO roadshow. Get developers building habits around Claude now, and those enterprise contracts look a lot more valuable when the bankers come calling.
Source: TechCrunch
SAUDI ARAMCO LEADS $800 MILLION BET ON OPEN-SOURCE AI: TOGETHER AI IS THE ANTI-OPENAI AND IT JUST BECAME AN $8.3 BILLION COMPANY
A company you have almost certainly never heard of just closed an $800 million funding round. That company is Together AI. It rents out Nvidia GPU clusters to developers who want to run open-source AI models. That is the whole business. And the world’s largest oil company just led the round.
Saudi Aramco’s venture arm stepping into AI infrastructure at this scale is one of those sentences that would have seemed like satire a few years back and now is just a Wednesday announcement. Alongside Aramco Ventures in the round: Vista Equity Partners, General Catalyst, Emergence Capital, Nvidia, and several others. Nvidia investing in a company that rents out Nvidia chips is a form of capitalism so circular it should be studied in business schools, but here we are.
The valuation trajectory tells the story. Together AI’s Series A in 2023 was $102.5 million. The Series B sixteen months ago was $305 million at a $3.3 billion valuation. Now the Series C is $800 million at $8.3 billion. The company has more than doubled its valuation in a year and a half.
What is driving this? The company says annual bookings just crossed $1.15 billion. It also claims that usage of open-source models across the industry has tripled in the past year. Those numbers are consistent with what you see on OpenRouter, the gateway platform that routes developer traffic across dozens of models. Open-source models are eating into the market share of closed frontier models faster than most of the industry expected.
The business model is clean. If you are a startup building on AI and you want to use Llama 4, Mistral, or any of the dozens of capable open-source models out there, you need compute to run them on. You can build your own GPU cluster, which requires capital and engineering bandwidth most companies do not have, or you can rent capacity from Together AI and pay by the token. For most developers, that decision takes about thirty seconds.
What makes Together AI specifically compelling is the customer list. Cursor, the AI coding tool that has taken over developer workflows, uses it. Cognition, which builds autonomous software engineers, uses it. Decagon, which handles AI customer service at enterprise scale, uses it. These are not experiments. These are production-grade AI applications under real load.
The bigger picture is what this round signals about where the money is going. Two years ago, everyone was racing to build the best foundation model. Now the money is chasing infrastructure. Compute, storage, networking, the pick-and-shovel businesses that win regardless of which model ends up on top. Together AI is the clearest expression of that thesis in the open-source ecosystem.
Also, Saudi Aramco is now a significant shareholder in an AI company that helps people avoid paying for closed models from American labs. The irony is rich enough to serve at dinner.
Source: TechCrunch
WHILE WASHINGTON WAS BUSY BANNING ITS OWN AI, CHINA BUILT A COMPETITOR: ZHIPU’S NEW MODEL IS CLOSING IN ON ANTHROPIC AND DEVELOPERS ARE NOTICING
For about three weeks in June, the US government did something genuinely unprecedented. It ordered Anthropic to pull its two most powerful AI models, Mythos and Fable, off the market entirely. No more access, no timeline for restoration, no public explanation beyond vague references to a jailbreak with national security implications. Companies that had built workflows and products around those models woke up to find them simply gone.
The export controls got lifted on June 30 and Anthropic started restoring access on July 1. Fine. But what happened during those three weeks matters more than the happy ending.
China’s AI industry took notes. Specifically, Zhipu AI shipped a model called GLM 5.2 while the American frontier was in regulatory limbo. And it landed with the kind of developer excitement that DeepSeek generated when it first showed up last year. The model sits within one percentage point of Anthropic’s Opus 4.8 on a closely watched agentic benchmark. It costs roughly one-fifth as much to run. And you can download it for free and run it on your own servers, which means no one can ever revoke your access to it.
Let that sit with you for a second. A Chinese AI company released a free, open-source model that is essentially competitive with the best Anthropic has to offer, during the exact window when Anthropic’s models were being pulled off the market by the US government.
Gabe Pereyra, co-founder of Harvey, the legal AI company that is one of the more sophisticated enterprise AI buyers in the world, told CNBC directly: “I’ve been consistently surprised by how quickly the open source has caught up. GLM 5.2, you’re seeing the first model where it’s really competitive with some of these closed-source frontier models.”
That is not a niche developer opinion from a Discord server. That is a serious enterprise buyer saying a Chinese open-source model is now in the same league as American frontier models.
What makes GLM 5.2 particularly interesting is that it is not just good at answering questions. It excels at planning, coding, testing, and looping, the exact kind of multi-step autonomous work that companies are racing to automate. That is the main battleground right now. That is where OpenAI, Anthropic, and Google are all fighting for enterprise contracts worth billions.
And here is the part that should make anyone in Washington uncomfortable. You cannot put an export restriction on a model that developers have already downloaded to their own servers in Frankfurt, Tokyo, or Singapore. Open-source AI is, by design, uncontrollable. The US government’s attempt to contain its own AI models had the practical effect of sending developers looking for alternatives. Some of those alternatives are Chinese. And once developers build habits around a model, they do not switch easily.
The lesson of those three weeks in June is not that export controls on AI are useless. It is that their cost and benefit are much more complicated than they look. Washington would do well to study what happened when it pressed pause on American AI and the rest of the world kept going.
Source: CNBC