Newsletter · July 5, 2026
Weekly Digest 27
Meta is reportedly building a cloud business to resell spare AI compute, which markets read as a threat to the neoclouds but which looks more like a large buyer turning contracted capacity into a monetizable asset; OpenAI floats giving the US government a 5% stake, testing whether frontier AI is too strategic for the state to stay only a regulator; a Ramp and Revelio study finds heavy AI adopters are hiring more rather than less, complicating the layoff narrative even as Goldman sees a monthly drag on job growth; and Etched raises $800M at a $5B valuation for Sohu, an ASIC that hard-wires transformer inference.
Topics we are tracking
Meta tests the resale market for AI compute
Source: Bloomberg — Meta Is Planning a Cloud Business to Sell AI Computing Power
According to reporting by Bloomberg, Meta is exploring a cloud business: a way to sell access to spare AI capacity, either through hosted Meta models or raw compute.
The market treated the story as a direct threat to the neoclouds. CoreWeave and Nebius sold off on the idea that one of their largest customers might become a competitor. The reaction is understandable, but it's more nuanced. Meta already has large forward capacity commitments with external providers, including CoreWeave and Nebius. Those agreements look much closer to infrastructure offtake than spot GPU rental. The exact termination rights are not public, but these are not contracts that disappear because Meta decides to test a resale product.
The better reading is that Meta has crossed the scale where compute becomes a balance-sheet asset. Meta has contracted more than 5GW of cloud and colocation capacity in the first half of the year alone, before counting its accelerating self-build activity. The company is buying enough compute that it now has multiple ways to monetize it.
The first use case is still frontier AI. Meta has not given up on training frontier models. Meta Superintelligence Labs remains the core reason for the buildout, and large-scale training still requires clean, homogeneous accelerator islands: one GPU generation, one memory profile, one network topology, one storage path, and one failure domain. Once a job is spread across thousands or tens of thousands of GPUs, the slowest or most failure-prone part of the system starts to dominate.
The second use case is Meta's core business: ads, ranking, and recommendations. This is the part of the story that is easy to understate. If larger recommendation models improve targeting, conversion, ad pricing, and time spent across Facebook and Instagram, compute can flow directly into the existing revenue engine. Meta may be able to scale ads recommendation compute by more than 10x as it moves to generative recommender systems (RecSys), which scale with compute — improving with both more training and more inference. That is a very different business from renting GPUs to startups.
The third use case is model distribution. Meta is reportedly in final talks with Anthropic for private Claude instances, similar in spirit to Bedrock, Foundry, or Vertex. This makes sense strategically. Meta has distribution, advertiser relationships, consumer surfaces, and internal demand for frontier model access. If it can serve its own models externally today and eventually offer third-party frontier models on Meta-controlled compute, the cloud business becomes less like bare-metal IaaS and more like token-as-a-service.
The fourth use case is the SpaceX-style deal: large-scale, short-duration compute sold at a premium to a frontier lab or enterprise buyer that cannot get enough capacity elsewhere. SpaceX effectively created a new market segment by selling large blocks of compute on unusually attractive economics, with short cancellation windows that preserve strategic flexibility. Meta does not need to become a broad neocloud with 30% gross margins. A few hundred megawatts of high-priced external compute could generate billions of dollars of revenue while preserving the option to pull capacity back into Meta Superintelligence Labs if internal progress improves.
SpaceX created a new market segment — short-term, large-scale compute at a premium
This is why the reported cloud idea is more plausible as an inference, token-serving, and overflow-capacity business than as a clean replacement for the neocloud training market. Meta's GPU inventory includes older A100 capacity, large H100 clusters, Blackwell systems, partner hardware, Meta's own MTIA silicon, colocation capacity, and contracted neocloud supply. That kind of fleet is harder to sell as pristine frontier-training infrastructure. It is easier to route as inference capacity. Different models, latency tiers, batch sizes, context lengths, and customer workloads can be assigned to different hardware pools. Older GPUs can still serve useful workloads. MTIA can absorb internal ranking and ads inference. H100 or Blackwell capacity can be reserved for higher-value serving or internal training.
H100 rental data show one-year contract pricing rising from the start of 2026, even after a long decline from 2024 levels. That does not mean every cluster is fully utilized. Meta can have idle pockets of compute while the broader market remains short of clean, externally available capacity. But this is not evidence that the compute shortage is over. If the market were already drowning in excess capacity, H100 rental pricing would likely be weakening, not firming.
Hyperscaler resale may create marginal supply risk for neoclouds, especially in inference. Idle H100 or Blackwell capacity will not stay idle forever if customers are willing to pay for it. But the core neocloud product remains different: clean accelerator blocks, power, networking, storage, tenancy, support, SLAs, and the ability to bring new capacity online faster than customers can build it themselves.
Large buyers will increasingly manage compute as an asset, not just an input. Some will resell underutilized capacity. Some will use it for frontier training. Some will push it into ads, recommendation systems, agents, and token-serving businesses. The important point is that compute demand remains large enough, lumpy enough, and urgent enough that every major owner of capacity now wants optionality.
OpenAI tests the sovereign wealth fund bargain
Source: TechCrunch — OpenAI proposed donating 5% of its equity to a US sovereign wealth fund
OpenAI has reportedly discussed giving the U.S. government a 5% stake. The proposal is still early, but it tests a larger thesis: frontier AI may be too strategic for the state to remain only a regulator.
Washington has already moved from light-touch oversight to direct industrial policy. Export controls determine who can access advanced chips. Safety reviews shape model releases. Procurement rules influence which companies become trusted suppliers. National-security pressure affects supply chains and customer access. The Trump administration has gone one step further, including direct equity stakes in strategic companies such as Intel. OpenAI's proposal would bring that same logic to frontier AI.
But "nationalisation" is too blunt a word. Different forms of public equity solve different problems. A direct distribution of shares to citizens tries to solve a legitimacy problem: ordinary people do not feel that they have agency in, or participation in, the upside of technological change. A government-held equity stake solves a different problem: giving the state a claim on a strategic asset.
The first version is mostly symbolic. A broad equity distribution would let the AI industry say that it is not building a future only for founders, employees, hyperscalers, and venture investors. It gives citizens a direct claim on the future being built around them. The problem is scale. A 5% stake in a $1 trillion company is worth $50 billion. Spread across roughly 130 million U.S. households, that is about $385 per household. Even if the stake doubled, it would still look more like a one-off stimulus check.
A government-held stake is more complicated. If the state owns part of a frontier lab, it becomes financially interested in the success of a specific company. That weakens policy neutrality. A new lab, an open-source challenger, or a foreign competitor would no longer be competing only against OpenAI. It would be competing against a company which the government has a direct interest in protecting.
It would also change how the company is perceived abroad. American AI already sits inside a geopolitical contest over chips, data, cloud infrastructure, and security guarantees. If frontier labs begin to look like extensions of U.S. state policy, foreign customers and governments will treat them differently. Palantir is a useful warning. The U.S. government does not own a stake in it, but its association with the national-security state already shapes how the company is perceived abroad.
Once the state owns part of a frontier lab, it may start using that ownership to shape content policy, usage monitoring, national-security workflows, model-release decisions, procurement access, and foreign deployment. At that point the lab starts to look less like a private company operating under regulation and more like a quasi-public utility. In the U.S., private companies acting under strong government pressure can sometimes start to look like state actors. That opens the door to due process, speech, administrative-law, and public-utility arguments.
This is the strongest objection to the government-equity model. The public does not necessarily feel richer because Treasury owns a stake in OpenAI. The legitimacy benefit is weak if the ownership is invisible to households. The costs however are much more concrete: conflicts of interest, regulatory capture, governance complexity, foreign mistrust, and the risk that frontier labs are pulled toward public-utility status.
There is a narrower and cleaner alternative: compensate the communities that actually host AI infrastructure. If a data center consumes local power, land, water, and grid capacity, the community should receive something tangible in return: lower power bills, direct payments, tax revenue, infrastructure funding, or long-term revenue sharing. This would also address the growing backlash against data centers in local communities.
The AI jobs story gets less clean
Source: Ramp — Does AI eliminate jobs? Economists find heavy adopters hire more
AI's impact on the labor market has become one of the most debated questions in Silicon Valley and beyond. The simplest version, often pushed by frontier lab CEOs, is that as companies adopt more AI, software will take on more cognitive work, substitute for employees, and eventually reduce headcount.
A new report from Ramp and Revelio Labs complicates that narrative. Using Ramp spending data linked to Revelio workforce records, the study looked at more than 21,500 U.S. firms and found that high-intensity AI adopters increased headcount by roughly 10% over the two years after adoption. Entry-level headcount rose even faster, by about 12%. But the firms spending most aggressively on AI are not representative of the entire economy. Ramp customers are more likely to be technical, growth-oriented, venture-backed, and already expanding. The report is therefore strongest as evidence against the crude layoff narrative, not as proof that AI mechanically creates jobs.
AI adoption sorts firms — heavy adopters pull ahead on headcount
- High-intensity adopters
- Low-intensity adopters
Goldman's new report on the AI Job Apocalypse lands in a similar place from the macro side. Goldman estimates that AI is already creating a 10,000 to 15,000 per month drag on U.S. job growth in the sectors where tools are being deployed first, including technology, consulting, and graphic design.
The report also assumes a 15% productivity uplift after full AI adoption and estimates that roughly 9% of U.S. workers, or about 15 million people, could be reallocated during the AI transition. Spread over a decade, that would still keep the annual unemployment impact below one percentage point.
The differentiator is adoption speed, which moves much more slowly than frontier model capability. A model being technically capable of a task does not mean the task gets automated. The system needs access to the right data, a reliable workflow, integration into existing processes, and a cost structure that makes deployment worthwhile. Large firms will automate before small firms, and only some exposed tasks will be economically attractive to replace.
AI can create a measurable drag in some exposed occupations while heavy AI adopters still hire more people overall. AI is replacing some tasks, reducing demand in some categories, and compressing some contractor markets. At the same time, firms that can absorb AI into workflows may expand because each employee becomes more productive.
The entry-level result is perhaps the most interesting part. The dominant narrative has been that junior white-collar work is the first layer to be automated: research, writing, analysis, customer support, basic coding, and operational coordination. Ramp/Revelio finds the opposite pattern among high-intensity adopters: entry-level headcount rose by about 12%. One explanation could be that AI makes junior workers productive earlier, while senior employees shift toward supervision, review, and workflow design.
We have long argued that AI adoption is not a clean labor-replacement shock but a sorting mechanism. Firms with the capital, data, management discipline, and technical culture to use AI heavily are pulling further ahead. They are hiring more because AI increases what those teams can do. Firms that only dabble do not show the same effect.
Etched raises for the inference bottleneck
Source: TechCrunch — Nvidia competitor Etched hits $5B valuation, $1B in sales for AI chip
Etched is the most interesting VC story of the week. The company says it has booked more than $1 billion in customer contracts for inference systems powered by Sohu, its transformer-specific AI chip, and has raised $800 million in total at a reported $5 billion valuation. Its first racks are expected to ship this summer, after TSMC manufactured the chip earlier this year.
Sohu is not a general-purpose GPU. It is an ASIC built specifically for transformer inference. GPUs spend a lot of silicon budget on flexibility: general-purpose compute, graphics heritage, broad software support, and hardware paths for many different workloads. That flexibility is Nvidia's moat. Etched is betting that, for inference, some of that flexibility has become overhead. If most high-value inference workloads keep looking like transformers, then a chip that hard-wires the transformer dataflow can spend more of its area, power, and memory bandwidth on the operations that actually matter. TechCrunch reported that Sohu is manufactured on TSMC's 4nm process and is designed only to run transformer models.
In transformer-based models the dominant workloads are: matrix multiplications, attention, feed-forward layers, KV-cache reads and writes, low-precision arithmetic, batching, and high-bandwidth memory movement. In inference, raw FLOPs are not enough. The system also has to move weights, activations, and KV cache through memory fast enough to keep the compute units busy. Etched's claim is that by removing support for non-transformer workloads, Sohu can deliver much better cost and power efficiency for the dominant serving workload. The company says it has designed the architecture to run math blocks at less than half the voltage of most AI chips, increasing FLOPs density.
Specialization only works if the workload remains stable. Sohu cannot be a great chip for every future model architecture. It is a bet that transformers, or transformer-like architectures, remain the center of gravity for language, image, video, agents, and reasoning inference long enough to amortize the chip, software stack, and systems buildout. If the market shifts toward very different architectures, or if Nvidia's Blackwell and Rubin roadmap compresses the cost-per-token gap, Etched's advantage narrows.
The reported investor base is also noteworthy. Jane Street, Hudson River Trading, Two Sigma, Stripes, Ribbit, VentureTech Alliance, and several high-profile AI researchers and founders are all involved. Jane Street and HRT stand out because they understand hardware advantage differently from ordinary venture investors. Both have built teams around low-latency FPGA and ASIC-based systems for trading, where latency, throughput, and infrastructure efficiency are not abstract engineering metrics but direct economic inputs. Their participation suggests that Etched is being underwritten as a bet that transformer inference becomes a market where specialized hardware can create a real cost and performance edge.
Seen on X
Other interesting stories
- Cosmos Institute: Can AI Make Scientific Breakthroughs?
- ChinaTalk: Rare Earths
- The Information: OpenAI discovers a new way to cut inference costs in half
- Reuters: Microsoft launches firm to help companies adopt AI with $2.5 billion
- Meta AI: From Brain Waves to Words — Brain2Qwerty offers a new path to communication without surgery
