AI Buildout Growth Rates

I continue to believe AI is the most revolutionary technology development of our time, and will have more significant impact on our lives and investments than the internet did. While I believe AI investment will grow in the years ahead, I believe we will continue to experience an “AI stock winter” that started in June. I remain invested in AI, and will continue, but I also remain committed to wide moat, owners of the architecture, or highly differentiated business , and have stayed away from memory, storage, or anything else that can be easily substituted by using a different vendor with a simple phone call. Examples in the first camp are Arista, ARM, hyperscalers, Nvidia, and TSMC. The latter camp would be neoclouds, Western Digital, Sandisk, and Micron. The latter have done very well over the last year. I have avoided these type of stocks, and will continue to do so, but I believe the only thing that has changed is to be even more diligent in doing so.

By my estimates, Micron has shipped 108 more the of DRAM bits over the last year than it did in 2023, but prices have gone up 345% over the same period. This shows that prices have a significant impact and prices are dependent upon supply and demand, which creates yet another variable to figure out, and it’s not a given that AI growth and memory/storage prices move in tandem. For example, Micron and Western Digital both crashed years before the 2000 tech peak.

Hyperscalers continue to spend more and more on AI buildout each year, and the rate of increase by year is impressive.

The issue I have is, they are reaching the capacity to continue their growth spending rates.

Hyperscalers are now having to tap debt and equity markets, and savings, rather than have their main business lines fund the rollout. Google has the added issue of the fact it can potentially cannibalize its bread and butter business as it does this.

COREWEAVE
CoreWeave is struggling with rising borrowing costs as it seeks billions of dollars to finance its AI data center expansion. Yields on some of its unsecured bonds have climbed to approximately 15%, prompting the company to pursue multibillion-dollar secured loans backed by Nvidia GPUs, computing equipment, and customer contracts at projected interest rates of 7–7.5%. CoreWeave is also expecting approximately $9 billion in customer deposits and prepayments to help fund expansion. Despite a $132.5 billion revenue backlog, the company faces substantial financing pressure from its $66.2 billion in debt and lease liabilities, $14.5 billion in quarterly capital expenditures, and $846 million in quarterly interest expenses. Its financing strategy increasingly depends on using customer commitments and infrastructure assets as collateral to obtain cheaper debt and sustain growth.

NEBIUS
Nebius (NBIS) is aggressively raising capital to finance its massive AI data center expansion, including $5.75 billion in convertible bonds, a $3.5 billion secured credit facility, and approximately $1 billion in stock sales during Q3 2026. The company expects to spend $27 billion on capital expenditures in 2026 and another $40–45 billion in 2027, despite projecting only $12–13 billion in revenue next year. Nebius plans to finance this expansion through customer prepayments, GPU-backed loans, convertible debt, and equity issuance, with a goal of reaching pretax breakeven by late 2027. Although it has secured relatively inexpensive financing compared with CoreWeave, its business model remains heavily dependent on continued lender support, access to capital markets, and customers fulfilling long-term contracts. The central risk is that Nebius is committing enormous amounts of capital years before its infrastructure can generate sufficient cash flow to sustain itself.

IREN
IREN (NASDAQ: IREN) is also relying heavily on external financing to fund its transition from Bitcoin mining to AI cloud infrastructure, but it has secured substantial funding on relatively favorable terms. The company arranged $3.65 billion in investment-grade GPU-backed financing at approximately 6% interest, supported by its Microsoft contract, with customer prepayments helping cover 96% of the associated $5.81 billion in GPU expenditures. It has also arranged approximately $2.8 billion in additional GPU financing, including a $2.4 billion facility at 9% interest, and had roughly $6.75 billion in convertible notes outstanding as of June 2026. In its November 5 earnings release, IREN reported $20.3 billion in contracted revenue backlog and said its target of more than $4 billion in annual recurring revenue was fully contracted and funded, although deployment has been delayed into early 2027. Like CoreWeave and Nebius, IREN remains dependent on debt, customer prepayments and potentially additional equity issuance to finance expansion well ahead of the cash flow it expects to generate. Its advantage is that major customer contracts have already helped it obtain substantial asset-backed financing at relatively attractive rates.

NSCALE
Nscale may be facing more serious financing pressure than its recent fundraising suggests. The British AI infrastructure company raised $2 billion in equity in March 2026, secured approximately $3 billion in GPU-backed loans in August, and announced $3.36 billion in pre-IPO convertible financing in September, including a $1 billion commitment from Nvidia. However, a November 13 regulatory filing revealed that Nscale’s debt surged from $137 million at the end of Q2 to $5.5 billion at the end of Q3, and the company warned it would need additional financing to meet its 2027 obligations, with no assurance that funding would be available on acceptable terms. Nvidia has reportedly considered guaranteeing billions in additional financing for Nscale’s West Virginia data center project supporting Anthropic. Despite reporting more than $103 billion in contracted business, Nscale remains dependent on continued borrowing, Nvidia’s financial support, and a planned IPO to finance construction. The central concern is that Nscale is accumulating debt rapidly before its infrastructure generates sufficient operating cash flow, while increasingly relying on its primary chip supplier to help finance the purchases.

There is growing evidence that Nscale, CoreWeave, Nebius and other AI infrastructure companies are having to work harder to secure financing. The change has become particularly noticeable in November 2026.

Reuters reported on November 9 that banks and credit investors are becoming increasingly reluctant to extend additional loans to AI infrastructure companies, even when major technology companies guarantee the financing. Lenders are scrutinizing customer contracts, demanding stronger collateral, and limiting their exposure to the sector.

Three developments stand out:

  • Nscale: Nvidia is reportedly considering guaranteeing billions in loans for Nscale’s West Virginia project. Nscale also disclosed that it needs additional financing for 2027, despite having already raised substantial capital.

  • Nebius: Reuters reports that Nebius explored issuing high-yield bonds in September but abandoned those plans in October, instead securing a $3.5 billion loan facility from Citigroup.

  • CoreWeave: Its declining bond prices and rising yields indicate investors are demanding greater compensation for lending to the company.

The Wall Street Journal also reported that lenders are becoming more selective and imposing stricter terms on AI infrastructure financing.

What I find particularly significant is that this isn’t necessarily a demand problem. It’s becoming a financing problem.

These companies have enormous customer commitments, but banks are increasingly questioning how much additional exposure they want to AI infrastructure.

That creates a potential turning point: AI computing demand could continue growing while capital spending slows because financing becomes too expensive or difficult to obtain.

For Nvidia and the semiconductor-equipment industry, that distinction is extremely important.

THE DEMAND IS STILL THERE. THE ABILITY TO FINANCE IT, ESPECIALLY GROWING AT THE RATES IT HAS BEEN, IS DIMINISHING.
Anthropic: Its latest IPO fact sheet shows Q3 revenue of $17 billion against $9.9 billion in compute and infrastructure costs. Its revenue-to-compute-cost ratio has improved substantially since early 2025, although it declined from Q2 to Q3. That’s encouraging evidence of improving economics, but not proof of sustainable free cash flow after all operating expenses and future commitments.

OpenAI: September’s approximately $50 billion annualized revenue is substantial, but the company faces the challenge of converting its enormous consumer and enterprise usage into enough cash to cover inference, training, and contracted infrastructure. Its reported revenue is also measured differently from Anthropic’s, complicating direct comparisons.

xAI/Grok: SpaceX IPO reporting showed xAI generated $3.2 billion in revenue in 2025 while recording approximately $6.4 billion in operating losses and spending $12.7 billion on capex. That illustrates how far its infrastructure investment was ahead of its operating earnings at that point.

Anthropic in particular, is giving the most clear financials.
Anthropic is demonstrating dramatically improving AI economics despite paying premium prices for short-term computing capacity. In Q3 2026, the company generated $17 billion in revenue against $9.9 billion in compute and infrastructure costs, leaving a $7.1 billion surplus before other operating expenses. Revenue grew 1,127% year over year, compared with a 382% increase in compute costs. Anthropic now generates $1.72 in revenue for every $1 spent on compute, up from just $0.40 in early 2025. Although this declined from $2.08 in Q2, partly because of approximately $5 billion in expensive short-term compute contracts, the company still reported $2.5 billion in adjusted operating profit, despite a $567 million GAAP operating loss. The trend suggests Anthropic is achieving substantial improvements in AI monetization and computing efficiency, even as capacity shortages temporarily pressure margins.

I plan on buying Anthropic stock when it goes public, assuming prices are OK. I am hoping the short term bearishness in AI allows for a more palatable opening, $1.3 trillion or so market valuation versus $2 trillion just a few months ago.

IS FRONTIER LLM A LOW MOAT BUSINESS?
This is quite possible. There are no real switching costs. Labs continue to one up one another. None are really pulling away. And META has entered the space with its Agentic AI. However, this is possibly similar to the early days of web search. Google dominates the market despite having little to no switching costs. It is quite possible we are in the early days of Alta Vista, Yahoo!, Lycos, Excite, and Google. I plan to play the Gorilla Game on LLMs.

NVIDIA
For some time, I saw Nvidia as the “one and done” investment for AI. I no longer believe that. Estimates are that Nvidia is used on some 70% of AI workload. Hyperscalers and even Frontier LLMS are developing their own ASICs that compete with Nvidia. The Gorilla Game was a post-hoc analysis of the 1990’s tech boom. We may never see it repeat again. We may never see another Intel again for the mere reason companies do not want to hand power to another company to create yet another Gorilla and strip themselves of that power. This is why every single hyperscaler and frontier lab is developing their own chips.

TSMC
This seems to have more of a lock on the AI chip buildout, but I could be wrong. Becuase regardless if they are using homegrown, Marvell or Broadcom, Nvidia, Intel, Cerebras, AMD, or anyone else, It winds up being built at TSMC.

CADENCE AND SYNOPSYS
I know these companies are incorporating AI, but I have no idea where AI displaces or reduces the demand for software. It is a difficult area for me to try and figure out. that goes not just for EDA software but for stuff like ServiceNow and DataDog. It is a real grey area for me. While Cadence and Synopsys should benefit from more companies developing their own chips, Anthropic is developing chips using AI instead of engineers.

This last point is one of the reasons that AI is such a disruptive force. We are not investing status quo with the benefit of a new growth area to tap into. We are experiencing a revolution that makes Salesforce eating Siebel Systems look like a grain of rice compared to the amount of eating there will be in the years ahead. Which is one of the reasons I plan on buying frontier labs when I can. I have not bought SpaceX yet.

CIRCULAR FINANCING
I am not sure what Nvidia will do regarding Circular Financing in the future.

I guess my point of all this is, I want/need to continue to invest in AI, but want to do so in a prudent manner, and expect it to be a bumpy road, not straight up.

PRICE MULTIPLES
If the rate of growth changes, it can have an impact on price multiples, especially for capital equipment makers like ASML, AMAT, KLAC, LRCX. These are mega caps now, and even though some of them are half off their recent highs, remain well over 2x historical price multiples. If growth rates moderate, they could fall further or become poor performers over the years ahead.

28 Likes

Thank you for sharing your thoughts!

What are your “AI” stocks that you believe are positioned better for the longer term? I’ve continued to hold for the past several years several companies which were doing well prior to the AI boom, and are also benefiting from the AI trade - CRWD, NET, AMD and SITM.

I have TSMC, NVDA (yes Nvidia but it’s no longer my main position), ARM, GOOG, AMZN, META, and MSFT as my AI core.

ARM: Market share for AI servers has gone from 15% to 60% over the past few years. I predicted this. I bought ARM at the IPO. Unfortunately, it hasn’t really impacted its revenues in the way I was expecting. However more and more AI processors are moving from X86 to ARM. It just does not seem to have the pricing power to really capitalize on it the way I would like. They are going to start making processors themselves, competing with their customers. According to a November 11, 2026 report in The Times, ARM founder Sir Robin Saxby made two significant points:

  1. AI is becoming dangerously speculative. He predicts a major market correction by the end of 2027, citing enormous infrastructure spending and questionable investment practices.
  2. He deliberately avoided manufacturing chips. His original strategy was to license ARM’s intellectual property rather than expose the company to the capital requirements and cyclical risks of selling semiconductors.
    ARM expects $15 billion in chip sales within 5 years.

Yes, I do have some CRWD, and NET, but not AMD and SITM. SITM is interesting. They seem to own the silicon based oscillators market. Even though there are competitors. Years ago I decided to never own semiconductors unless there was a programming aspect. NVDA and ARM both fit this. SITM does not. But they are dominant in a niche. Alterra, Lattice and other FPGA are also programmed and while they do not have a proprietary programming language, they are designed in and cannot be easily switched out. SITM, I understand they often work in unique packages so they cannot be easily switched out, unless a company specifically decided they were going to target that specific device, and I think right now it’s not on the radars of enough larger players. Altera is supposed to be spun off from Intel in the near future.

18 Likes