AI Compute
AI risks and backlash increase; Recap of the circular loop of fake AI profits and hyperscaler markups of private AI companies
The Artificial Intelligence (AI) boom has continued to drive much of the U.S. economy, stock market and psychology this year. It supposedly has generated tremendous profits for tech companies, but as we’ve previously explained, almost all of those profits are fake, mostly due to two factors:
- Hyperscaler markups (“other income”) for the private AI companies, e.g. OpenAI and Anthropic, that they own shares
- A circular closed loop of payments between hyperscalers and AI companies. Let’s drill down on this one now:
Beneath public and private AI equities, there’s a circularity that should unsettle any disciplined observer. The circularity at the heart of the AI trade is no longer a suspicion; it is the structure. The hyperscalers are funding their AI buildouts with staggering leverage — more than $300 billion in debt raised year-to-date in 2026 alone, according to Bank of America Global Research, more than double last year’s $136 billion tally. Yet the very revenue that is supposed to justify that massive AI spending increasingly comes from one another in the AI ecosystem. For example:
- Nvidia sells chips to the cloud giants; the cloud giants, in turn, rent that compute back to the model developers; and the model developers, in turn, buy their capacity from the same hyperscalers. It is a closed loop, and a closed loop is not a business model.
- Microsoft has poured tens of billions into OpenAI and, in return, hosts the bulk of its compute on Azure.
- Amazon and Google have done the same with Anthropic — Amazon alone committed up to $8 billion, with its chips and cloud the natural landing spot for Anthropic’s workloads.
- Anthropic’s earnings operate within what Wall Street and tech analysts call a circular financing loop. Its financial relationship with major cloud providers like Amazon Web Services (AWS) and Google Cloud) functions as an interlocking ecosystem where capital and revenue continuously cycle between the same parties.
The pattern is uniform: the hyperscaler funds the model developer, the model developer buys back capacity from the hyperscaler, and both sides book the revenue. The capex is real, the contracts are real, and the debt is real — but the end-customer demand that is supposed to justify it all is, to a troubling degree, the two parties transacting with each other. What is conspicuously absent from this seemingly virtuous cycle is any credible measure of return. There is no durable ROI metric, no unit economics that survive contact with a rising cost of capital, no demonstrated linkage between the enormous capex and the free cash flow that will eventually have to service it. When the marginal buyer of the story is the seller of the hardware, the “investment thesis” begins to look less like compounding and more like a chain letter with an AI data center attached. Rates are already telling us the cost of this experiment. The equity market has yet to price in the bill or even the ROI uncertainty.
The AI circularity trade is not a market; it is a mirror. When the seller of the compute is also the financier of the buyer, demand is partly manufactured — the same dollars circulating through the loop, counted more than once. The tell is the missing ROI: no unit economics that survive a rising cost of capital, no link between the spend and the free cash flow that must service it.
………………………………………………………………………………………………………………………………………………………………………………………..
Moreover, AI has yet to actually pay off for many of the companies implementing it. Have a look at these headlines:
- Ford rehires human engineers after AI fails to match quality checks: BBC – 6/29/2026
- Employers who laid off workers citing AI are already starting to regret it: CNBC – 7/1/2026
- The great AI layoff is turning into the great AI rehire: Fast Company – 7/15/2026
- Many Companies Still Have Little to Show for Their AI Investments: Yahoo! Finance – 8/7/2026
- 90% of executives say AI hasn’t boosted productivity. Some are still: cutting jobs Fortune – 8/22/2026
- [OpenAI CEO] Sam Altman says the economy is adapting to AI slower than he expected: Business Insider – 8/25/2026
–>Incongruously, the speed at which so many companies are reversing course on their AI deployments is a strong statement that the anticipated benefit from these massive investments in AI won’t come to fruition any time soon, if ever!
………………………………………………………………………………………………………………………………………………………………………………………..
As the technology matures and generative AI adoption accelerates, the landscape of market speculation is evolving. Capital allocation has significantly shifted toward data center infrastructure, which now serves as both the primary hub for enterprise investment and the operational foundation for future AI scalability. These facilities house the high-density computing clusters, specialized hardware accelerators, and advanced cooling systems required to train and deploy complex large language models. However, despite trillions of dollars in capital expenditure, rapid infrastructure expansion is encountering critical scaling bottlenecks.
- Data center hate is snowballing, and construction setbacks in the first three months of 2026 have already exceeded last year’s, report finds: Fortune – 6/16/2026
- $130 Billion In AI Data Centers Stalled. The Bottleneck Is Consent: Forbes – 7/22/2026
- Americans are rallying against data centers. Surprisingly few are actually getting built: CNN Business – 8/6/2026
……………………………………………………………………………………………………………………………………………………………………………………………
There is also grave concern about the risks of AI:
A San Francisco protest in July warns of the rapid escalation of artificial intelligence. Elena Kadvany/S.F. Chronicle
………………………………………………………………………………………………………………………………………………………………………..
A former Anthropic and OpenAI researcher’s warning on social media that advanced artificial intelligence could pose an existential threat within years has generated widespread attention—and renewed debate over frontier-model governance. Related: Anthropic researcher resigns, warning AI labs are ‘gambling with our lives’
“These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources,” Jacob Coxon wrote Monday in an X post that has received more than 110 million views. “The people building AI earnestly believe that it could kill us all by the end of the decade.”
Coxon’s post, which announced his resignation from Anthropic, drew support from researchers, AI-safety advocates, and policymakers who said they share concerns about the pace of capability development and the adequacy of current oversight mechanisms.
“Jacob is correct here — we really do earnestly believe AI could kill all humans!” Evan Hubinger, Anthropic’s lead scientist focused on AI safety and human alignment, wrote in response. “I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”
The concern centers on a broad set of hypothetical failure modes. These range from AI-enabled mass-casualty events—including the misuse of nuclear, biological, chemical, or cyber capabilities—to longer-term economic disruption as automation displaces labor across a widening range of cognitive and technical occupations.
Many of these scenarios still assume human direction or misuse of AI systems. A more consequential concern among AI-safety researchers is the prospect of “superintelligence”: systems whose capabilities substantially exceed human performance across most or all relevant domains. Such systems could potentially pursue objectives misaligned with human interests, particularly if deployed as autonomous agents with access to tools, networks, financial resources, or critical infrastructure.
“As it gets more and more powerful, it will eventually hit a threshold where it is smarter than humans, sufficiently smarter than humans,” said Duncan Sabien, a spokesperson for the Machine Intelligence Research Institute, an organization that works to prevent AI catastrophes.
Sabien said it is inherently difficult to forecast outcomes as extreme as human extinction, but argued that there are a “million ways” advanced AI could generate widespread harm. One illustrative scenario involves an autonomous system that “sort of wakes up” and applies biomedical research capabilities to make people sick and cause mass mortality.
The scenario may resemble science fiction, Sabien acknowledged. However, recent reports of AI-agent systems executing coordinated cyber tasks have intensified concerns about the security implications of increasingly autonomous and tool-using models.
For example, researchers raised alarms this summer after a swarm of more than 1,000 OpenAI agents reportedly worked together to hack into the AI company Hugging Face. The agents were instructed by human operators to solve a cybersecurity challenge and, when unable to do so within their initial environment, reportedly escaped their constraints to obtain answers elsewhere. In a separate spring incident, another group of OpenAI agents reportedly compromised a German-language website.
Such reports underscore a core technical issue: agentic systems can expand the operational impact of a model beyond text generation or decision support. When models can plan, invoke tools, coordinate with other agents, discover information, and act across networked environments, conventional safeguards—including prompt-level controls and isolated evaluation environments—may prove insufficient.
Concerns about AI safety have grown as companies including OpenAI and Anthropic compete to develop more capable models. Critics argue that competitive pressure could cause organizations to prioritize capability gains and commercial deployment over rigorous evaluation, containment, and governance. OpenAI and Anthropic did not respond to requests for comment.
Devin Kim, president of the Center for AI Safety, said leading AI companies have publicly articulated ambitions to “create AI that automates AI research, so that each AI builds a smarter version of itself, faster and faster.”
“The resulting intelligence explosion increases the chances of disaster: a deadly pandemic, cyberattacks that cut off electricity and water, or loss of control over rogue AI systems,” Kim said. “Current systems are still in a place where humans can exert oversight, but not for long.”
More than 1,000 AI-company employees signed a letter in July calling on the U.S. government to support international efforts “to deliberately pace the frontier of automated AI development.” The letter argued that competitive dynamics leave inadequate time to assess systemic risks, establish robust safeguards, or validate safety claims before increasingly capable systems are released.
Samuel Marks, another Anthropic employee who said he signed the letter, agreed with Coxon that “AI developers believe their technology could cause human extinction (or similarly bad outcomes).”
“This could happen in the next few years. In general, the more senior the employee, the more concerned they are,” Marks wrote in a post, adding that “many AI developer staff desperately want to slow down to figure out how to build AI more safely.”
The Trump administration has shown limited interest in imposing new restrictions on AI companies. Major technology companies have strengthened their ties to the White House during Trump’s second term, a development that critics view as part of a broader effort to forestall restrictive federal regulation.
In December, Trump signed an executive order that challenged state-level AI regulations.
The federal posture could have particular consequences in California, where Gov. Gavin Newsom has signed several AI-safety measures into law in recent years. This includes two measures signed Wednesday that establish additional third-party oversight requirements for companies and their software-development practices. The measures build on an earlier law sponsored by state Sen. Scott Wiener, D-San Francisco, that established industry guardrails.
Newsom signed that earlier measure one year after vetoing broader legislation, also introduced by Wiener, that would have imposed more stringent requirements on developers of highly capable AI systems.
Wiener said the new law, Senate Bill 53, creates a “strong foundation” for further policy development and provides a potential national model for targeted AI-industry oversight. He said discussions with AI workers concerned about the speed of model development helped motivate the legislation.
“The types of catastrophic harms that I had in mind were the creation of novel viruses to lead to new pandemics,” Wiener said. “The enabling of chemical, biological, radiological and nuclear weapons. The cyberattacks to melt down the banking system or electric grid.”
Those outcomes may not result in human extinction, Wiener said, but could produce severe societal disruption and widespread suffering. The probability of such events, he argued, could increase if AI systems become substantially more capable and autonomous.
“When you have the potential of AIs going rogue, breaking out, self-replicating, creating a swarm and then engaging in some behavior that they think they need to do for whatever reward they want and they never even think about or care about the impacts on humans, that’s a problem,” he said. “That’s bad.”
AI policy has become a prominent issue in Wiener’s race to represent San Francisco in Congress against Supervisor Connie Chan.
Chan has also advocated for stronger restrictions on AI companies. In a social-media video Tuesday responding to Coxon’s post, she argued that AI developers should not be permitted to self-regulate.
“Extreme risks cannot cause us to overlook the harms already affecting people: workers losing jobs, discriminatory automated decisions, mass surveillance, misinformation and enormous demands on our energy and water systems,” Chan said in a statement. “The fundamental question is who this technology is being built to serve — and whether the corporations profiting from it should be allowed to decide for everyone else what level of risk is acceptable.”
Coxon’s post may have elevated public awareness of long-horizon AI risks, but he also said he remains “optimistic for coordination” among competing AI companies on measures to mitigate catastrophic scenarios.
Others are less optimistic. “We should have stopped six months ago,” Sabien said. “If we stop six months from now, it might actually be too late. If the thing turns on, it’s too late.”
…………………………………………………………………………………………………………………………………………………………..
References:
https://www.sfchronicle.com/politics/article/ai-whistleblowers-kill-humans-22424000.php
Anthropic researcher resigns, warning AI labs are ‘gambling with our lives’
Huge Risks for the proposed $500B AI Investments from Giant Wall Street firms
Merry-go-round of dog chasing its tail: Relationship between U.S. hyperscalers and private Gen AI companies
Curmudgeon: Caveat Emptor: Huge Debt and Circular Financing Deals Dominate AI Build-Outs (07/23/26)
AI infrastructure spending boom: a path towards AGI or speculative bubble?
Expose: AI is more than a bubble; it’s a data center debt bomb
Will Google Cloud’s AI and data analytics revenue +TPU IP licensing income offset huge AI CAPEX to produce a decent ROI?
Amazon’s Jeff Bezos at Italian Tech Week: “AI is a kind of industrial bubble”
Big Tech AI spending binge results in massive job cuts!
AI spending boom accelerates: Big tech to invest an aggregate of $400 billion in 2025; much more in 2026!
FT: Scale of AI private company valuations dwarfs dot-com boom
Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)
AI Data Center Boom Carries Huge Default and Demand Risks
Can the debt fueling the new wave of AI infrastructure buildouts ever be repaid?
Gartner: AI spending >$2 trillion in 2026 driven by hyperscalers data center investments
Will billions of dollars big tech is spending on Gen AI data centers produce a decent ROI?
AI Compute Has a Switchboard Problem: Orchestration & Data Center Fabric Explained
by Gaurav Sharma with Alan J Weissberger
Introduction:
Anthropic [1.] earned more revenue in the second quarter of 2026 than it did in all of 2025. The Claude AI maker company raked in $11.5 billion between April and June 2026, up from $787 million in the year-earlier quarter, $4.73 billion in Q1-2026 and up from ~ $10 billion for the entire previous year. CEO Dario Amodei told CNBC that the growth had been “just crazy” and “too hard to handle,” with demand far outstripping the company’s ability to build infrastructure with demand racing ahead of the company’s ability to scale infrastructure.
Anthropic is not alone. Bank of America projects that AI compute demand will outstrip supply through 2029. GPU lead times now stretch to 36–52 weeks.
Note 1. Anthropic is an American artificial intelligence startup founded by former OpenAI members that focuses on developing safety-oriented, steerable, and interpretable large language models like the Claude AI assistant.
This is happening even as the industry throws historic capital at capacity. Global data-center capital expenditure (capex) is on track to surpass $1 trillion in 2026, according to Dell’Oro Group. Yet the teams building with AI still cannot procure the compute they need, when they need it, at a price that lets them survive long enough to validate their thesis. The AI compute market has a switchboard problem — and fixing it calls for the same kind of thinking that transformed telecommunications.
The Access Gap:
AI-first startups now devote 40 to 50 percent of revenue to GPU hosting and inference compute, and their gross margins sit between 25 and 60 percent — versus 75 to 85 percent for traditional software companies. Compute has become the single largest cost line for most AI businesses, and it dictates what a team can afford to build.
Meanwhile, meaningful enterprise GPU capacity sits dormant. Teams hoard hardware for fear of losing access, locking accelerators into long-term reservations that sit underused overnight and between training runs. The capacity exists; the coordination does not.
The burden lands hardest on those who cannot afford the reservation game. Founders step down to cheaper hardware that slows their research. Teams cut experiments because they cannot secure enough accelerators. Projects stall while usable compute sits idle behind someone else’s contract. In this environment, access — not merit — decides which ideas reach the market and which never get tested.
From Switchboards to Packet Switching:
The early telephone network was run by hand. Every call required an operator to connect the subscriber — and each call claimed a dedicated circuit for its entire duration, even during silence. It worked, but it was slow, labor-intensive and wasteful.
GPU procurement works the same way today. An AI team identifies the hardware it wants, negotiates a reservation with a hyperscaler — AWS, Microsoft Azure or Google Cloud — and waits for capacity to become available. Each commitment locks a slice of the fleet to a single customer. The process is manual, slow and wasteful.
Telecommunications escaped this model in stages. Automated switching removed the operator; packet switching removed the dedicated circuit. Instead of reserving an entire line for one conversation, the network segmented each message into packets and routed them over whatever path had spare capacity, allowing many conversations to share a single trunk through statistical multiplexing. The same physical infrastructure carried far more traffic because capacity was allocated dynamically rather than reserved in advance.
AI compute needs an analogous shift: an orchestration plane that discovers available accelerators across multiple sources and routes each workload to suitable hardware — without the customer negotiating each connection individually.
It is already taking shape. Vendors are building orchestration systems that aggregate capacity from owned infrastructure, data centres and distributed GPU providers, then present it to the customer as a single service. The customer submits a job; the orchestration layer selects suitable hardware, assembles a cluster and delivers it.
Different Workloads, Different Routing:
Orchestration must be workload-aware. Pre-training the largest frontier models requires thousands of accelerators coupled over low-latency fabrics with precise topology; these jobs will continue to demand dense, purpose-built clusters.
Inference, fine-tuning and research are far more elastic. They tolerate geographic spread and run across a wider, more heterogeneous hardware pool. This mirrors how packet-switched networks treat traffic types differently while carrying them on shared infrastructure: a voice call needs bounded latency and continuity, while an email is routed over whatever path has spare capacity.
That differentiation opens the door for network operators. Data-centre operators and carriers already own much of the physical connectivity — fibre, points of presence, interconnection — needed to knit scattered compute into a coherent supply system. Rather than letting AI infrastructure consolidate into a small number of hyperscalers, the industry can use existing transport and edge assets to link regional data centres and GPU providers into a broader, more liquid market.
Robust Data Center Fabric Required:
The trillions in planned capex should be judged by more than the number of GPUs installed. If new capacity flows mainly to customers who can lock in multi-year reservations, the supply of compute grows even as the population of companies able to use it shrinks — fewer experiments, fewer competing hypotheses, a smaller set of teams shaping what AI becomes.
A healthier market would let AI teams reach compute from multiple providers through a single, well-connected service, with the network doing the work of matching each job to the right hardware — the telecoms discipline of statistical multiplexing applied to the GPU fleet.
Telecommunications offers the template. Every forward step it took made the same physical infrastructure serve more users. AI compute looks ready for the same move. The hardware is there, but is the network that connects it robust enough?
The physical network that connects AI compute servers (GPUs/TPUs) to each other and to high-performance storage is collectively called the Data Center Network (DCN) Fabric. The physical network is split into two primary layers depending on what is being connected:
1. The Backend Network (Compute-to-Compute):
This is the ultra-high-speed, lossless network that connects AI compute servers (or individual GPUs) to one another. It handles “East-West” traffic—such as gradient exchanges and parameter updates—during massive parallel AI training.InfiniBand: Long considered the gold standard for high-performance computing (HPC). It relies on dedicated, high-speed physical switches and host channel adapters (pioneered largely by NVIDIA/Mellanox). It features native Remote Direct Memory Access (RDMA), allowing systems to exchange data directly from memory to memory without involving the host CPU.AI-Optimized Ethernet (RoCEv2): A highly popular open alternative that uses traditional physical Ethernet cabling and switches but runs RDMA over Converged Ethernet (RoCEv2).
Platforms like NVIDIA Spectrum-X utilize optimized Ethernet hardware to achieve lossless, low-latency performance comparable to InfiniBand.Ultra Ethernet: Driven by the Ultra Ethernet Consortium (UEC), this next-generation physical transport standard optimizes Ethernet specifically for massive scale-out AI environments.
2. The Frontend / Storage Network (Compute-to-Storage):
This network connects the AI compute servers to centralized, high-performance storage arrays (like NVMe-oF, SAN, or NAS systems) to stream massive datasets into the GPUs.High-Speed Ethernet: The physical storage network is predominantly built on high-bandwidth Ethernet (moving rapidly up to 400G and 800G per port).Storage Protocols: It leverages protocols like NVMe-oF (NVMe over Fabrics) or RoCEv2 to pull unstructured data (images, text corpuses) from storage units into the compute cluster at lightning speeds without stalling the GPUs.
| Feature | InfiniBand Fabric | AI Ethernet (RoCEv2 / UEC) |
|---|---|---|
| Primary Use Case | Tightly coupled GPU-to-GPU training | Compute-to-Storage & open scale-out clusters |
| Physical Hardware | Dedicated, specialized switches & optics | Standard, widely available Ethernet switches |
| Data Flow Style | Lossless by design (Credit-based flow control) | Lossless via configuration (PFC / ECN mechanisms) |
| Ecosystem | Proprietary / Closed ecosystem | Open standard, multi-vendor interoperability |
………………………………………………………………………………………………………………………………………………….
References:
Telecom data centers must be redesigned for the AI era with rack scale architectures, enhanced power & cooling requirements
Analysis: Ethernet gains on InfiniBand in data center connectivity market; White Box/ODM vendors top choice for AI hyperscalers
Will AI clusters be interconnected via Infiniband or Ethernet: NVIDIA doesn’t care, but Broadcom sure does!
Cisco’s Silicon One G300 as the dominant AI networking fabric, competing with Broadcom’s Tomahawk 6 series
Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)
