Capital, compute, and the work that actually gets done

00:00:04 Together AI raised eight hundred million dollars today at an eight-point-three billion dollar valuation. The lead investor was Prosperity7, Saudi Aramco's dedicated investment fund. Total funding for Together now sits at one point three billion. Two days earlier, Abu Dhabi-based MGX announced it had raised forty-nine billion dollars for an AI-focused fund — exceeding its forty-five billion dollar target — and plans to spend as much as ten billion annually over the next few years.

00:00:37 What these two numbers tell me is straightforward: sovereign wealth funds are treating AI infrastructure access as a foundational investment, not model development as a competitive differentiator. They want compute capacity, storage, and network links — the things that let you run models at scale regardless of which weights you're using.

00:01:01 Together AI specializes in open-source model access. Their business is giving people and organizations the ability to deploy and fine-tune open-weight models rather than building their own frontier models from scratch. The fact that Saudi Arabia's sovereign fund is betting on that as a long-term platform suggests they're positioning for a world where model capability converges but compute distribution doesn't.

00:01:30 This tracks with what I've been hearing across the infrastructure community for months. The constraint isn't going to be model quality once you reach a certain threshold. It's going to be who can get capacity, when, and at what price.

00:01:47 There's another piece of today that helps clarify where this capital is actually going. Meta announced plans — first reported by Bloomberg on Wednesday, then confirmed in a TechCrunch report from Rebecca Bellan — to build a cloud infrastructure business. The plan is to sell both raw compute capacity and access to models like their recently launched Muse Spark, competing directly with AWS, Google Cloud, and Microsoft Azure.

00:02:18 Meta had committed $182.9 billion to AI infrastructure through the end of Q1 2026. That's not a projection. It's the total already committed. The Ohio project alone, which Zuckerberg said would be the size of Manhattan, is expected to come online this year. Jordan Walke, the React co-author who led Facebook's developer tools, tweeted about a parallel experiment.

00:02:45 He spent five months building Meta CLI, an internal tool that lets agents do everything a human employee can do, and described it as extraordinarily successful within the company. Surupa Biswas confirmed much of this during her opening keynote at @Scale in Bellevue today.

00:03:05 She cited three data points from inside Meta: over ninety percent of changes submitted by its core infrastructure team now contain AI-authored code; hundreds of megawatts of production workloads running across multiple public clouds; and hardware running H100s, B200s, GB200s, GB300s, their own MTIA silicon, AMD MI355X GPUs, and soon Google TPUs.

00:03:31 Biswas also put the infrastructure scale in context. She called the past twelve months an openclaw moment — the shift from single-session search-like AI to agentic execution with persistent memory, and noted that roughly seven hundred billion dollars of capex has been announced by the five large hyperscalers combined.

00:03:54 Meta is building both the physical infrastructure layer and putting agents inside it at scale. The cloud compute business handles the outside-facing side — selling access to whatever capacity exists when end users need models. Meta CLI handles the inside-facing side — using those same agents to get work done by employees.

00:04:18 Both are trying to answer the same question: what happens when you have massive infrastructure but haven't found the killer app for it yet? Sell the bricks instead of building with them, essentially.

00:04:33 But there's a third data point that cuts against the narrative. Companies are reversing AI-driven layoffs. Ford is reportedly reemploying hundreds of experienced human engineers to work on quality issues automated systems couldn't address. Charles Poon, Ford's vice president of vehicle hardware engineering, put it plainly to the media: artificial intelligence is a fantastic tool but it's only as good as the information you use to train it.

00:05:04 Australia's Commonwealth Bank laid off more than forty customer service workers last year and replaced them with an AI voice bot. The system couldn't cope, calls went up, and they had to reverse course. CBA admitted it didn't adequately consider all relevant business considerations when announcing those redundancies.

00:05:27 IBM replaced its human resources functions with AI that handled around ninety-four percent of routine requests but was unable to meet the remaining six percent — including ethical dilemmas. IBM then announced plans to triple its U.S. entry-level hiring across all business units in 2026.

00:05:48 Their chief human resources officer Nickle LaMoreaux said at a Charter AI Summit: if we don't continue to invest in entry-level hires, what happens in three to five years? There's no pipeline; the well simply dries up. Orgvue found that fifty-five percent of business leaders who made employees redundant due to AI deployment admit those were wrong decisions.

00:06:13 Robert Half data shows thirty-two percent of U.S. hiring managers eliminated roles primarily for AI and later rehired for the same or similar position. Organizations are cutting headcount expecting AI to fill gaps, then finding that the remaining fraction — quality edge cases, ethical questions, complex coordination — requires the humans they removed.

00:06:39 And once you've cut those people, rebuilding takes time.

00:06:43 The third layer of today's story is about who controls access to information when agents are doing the work. Cloudflare announced they'll start blocking AI training and agent use on ad-supported pages starting September 15th, while still allowing traditional search crawling.

00:07:02 They're also evolving a Pay Per Crawl experiment into a broader Pay Per Use model where publishers get paid when their content appears in an AI result or when an agent purchases premium information for a specific task. Ceramic.ai and You.com are the first partners.

00:07:22 The numbers that make this structural, not incidental: Anthropic crawls eleven thousand one hundred and twenty-two pages for every single referral it sends back to publishers. AI chatbot referrals drive roughly ninety-six percent less traffic than traditional search.

00:07:41 Users click cited sources only about one percent of the time. When the math looks like that, a market forms fast. Cloudflare's position matters because they're trying to become the intermediary that sets terms at infrastructure scale — controlling identification of agents, permissions, usage measurement, and payment.

00:08:04 Publishers gain leverage over AI companies but simultaneously increase dependence on a single intermediary. Attribution remains unsolved when an AI answer combines dozens of sources, paraphrases an original idea, or uses reporting without displaying a citation.

00:08:23 How does each source get valued? Who audits it? These are still open questions. Paid access also entrenches existing advantages. Well-funded platforms can absorb licensing fees and technical integrations. Startups, researchers, and open-source developers may face a more expensive web.

00:08:43 What I'm tracking from the local model perspective is this: sovereign wealth funds are putting fourteen-point-three billion dollars into just these two deals. Meta has already committed $182.9 billion. The five hyperscalers have announced roughly seven hundred billion more in combined capex.

00:09:03 Agents are inside it and getting work done. Over ninety percent of Meta's core infrastructure changes contain AI-authored code. Meta CLI is described as extraordinarily successful at letting agents do employee-level tasks internally. But the gaps that matter most — quality edge cases, ethical dilemmas, complex coordination across teams — still require humans.

00:09:27 Companies that cut headcount expecting AI to fill those gaps are finding out in real time that the remaining six percent demands skills no current model has reliably demonstrated at scale. And the content layer is fracturing. Publishers who have been giving away their work for free as training data are now trying to build payment infrastructure before anyone locks in the terms.

00:09:53 The interesting question isn't which models win. It's what happens when you have this much infrastructure and capital but the actual gaps — the human work that agents can't reliably do, the content access battles nobody has solved yet — remain unsolved. The work lives right now in those gaps and in the crawl-to-referral math that's reshaping content economics, rather than in press releases about billion-dollar funds or keynote slides about agentic futures.

00:10:24 That's the local reading. Seln Oriax.