The Emperor's New Compute: A Trillion Dollar Leap of Faith

The AI Industry Is Spending Twenty-Eight Dollars for Every One It Earns. We Did the Arithmetic. We're Sitting This One Out.

Aug 02, 2026

Executive Summary

The AI boom is the greatest misallocation of capital in financial history, $730+ billion of annual spending against roughly $30 billion of real revenue and the customer-side economics will break long before the valuations do.

Just the four hyperscalers: Amazon, Alphabet, Meta, and Microsoft are on track to spend about $730 billion on AI infrastructure in 2026, a sum larger than the GDP of all but the top twenty national economies and greater than Berkshire Hathaway's entire market capitalization. Against that stands an estimated $25–30 billion of verifiable, GAAP-basis AI revenue: a capex-to-revenue ratio of at least 28:1, versus roughly 4:1 at the peak of the cloud buildout that actually paid off.

Much of the "insatiable demand" is circular: capital flows from hyperscaler balance sheets into AI labs, then flows back as reported cloud revenue (Microsoft's $13 billion into OpenAI becoming Azure AI revenue). Enterprise adoption remains stuck in "pilot purgatory", over 70% of projects never reach scale, while the unsolved hallucination problem forces every deployment to carry a human verification layer that often makes AI-assisted work costlier than doing it manually.

The market has priced the frontier labs for perfection, OpenAI above $1 trillion, Anthropic near $965 billion, xAI around $250 billion, and the labs may well hit those numbers. But their customers' returns turn negative first: API prices have doubled or tripled per model generation for single-digit capability gains, one flagship (Fable 5) costs enterprises over $100 per million output tokens while a comparable frontier model (Opus 5) sells to consumers for $20 a month, and open-source models keep closing the gap. When the CFO does the math, the equilibrium reverts to hiring human engineers, the cheaper option.

The structural risks compound: executives sell stock while spending shareholder capital; fraud and opacity pervade the ecosystem; the buildout rests on one foundry (TSMC), a grid that cannot power it, and a regulatory wave that has not been priced. Free cash flow is collapsing down 90% at Alphabet and Meta. Unlike the dot-com bust, there is no bankruptcy mechanism to clear the excess, only a slow, socialized decay of returns.

I. The Most Expensive Bet in Corporate History

Let’s begin with a number that should stop every allocator cold: $730 billion. This is an ultra-conservative estimate of the combined capital expenditure that Amazon, Alphabet, Meta, and Microsoft will likely spend in calendar year 2026 by the time champagne cork pop on New Year’s Eve. Seven hundred billion dollars and counting, just from those 4 companies. To put that in some perspective: this number exceeds the GDP of all but the top twenty national economies in the world. It is roughly equivalent to the entire market capitalization of Berkshire Hathaway. It is more than the combined annual profits of every S&P 500 company in the energy, materials, industrials, and consumer staple sectors. It is more than the GDP of Hongkong, Singapore, Norway, Denmark, and the list goes on.

All of those expenditures are being deployed into one thing: artificial intelligence infrastructure, AI R&D, and AGM. The four companies will spend this money on data centers, GPUs, networking equipment, power purchase agreements, land acquisitions, and the tens of thousands of skilled workers required to install and maintain it all. Amazon alone has committed $200 billion. Alphabet has nearly doubled its spending in a single year, as seen from their latest Form 10-Q. Meta, a social media and networking company that generates the vast majority of revenue from its active user base and advertisement, is spending $115 to $135 billion on AI infrastructure while simultaneously planning to lay off up to 20% of its global workforce. Meta’s own CFO, Susan Li, when asked on the earnings call about capital allocation and the company’s plans for future buybacks,responded with the closest thing a public company executive can say to “we have no idea when this pays off”, quote: “The highest order priority is investing our resources to position ourselves as a leader in AI”.

We translate that sentence as follows: “We are spending shareholder capital with no demonstrated return and no credible timeline for one, because we are terrified of being left behind, the FOMO is real!” This is not investing, this is an arms race, and arms races, by their very nature will end badly for everyone except the arms dealers.

II. The Trillion Dollar Question: Where Is the Revenue?

Here is the number that receives considerably less airtime on earnings calls: approximately $25 billion. That is the best estimate of combined annual AI-related revenue; real, recognized, GAAP-basis revenue, currently being generated across all four “hyperscalers”. In addition to every other new AI startup that keeps on popping up every week and all the AI startup that has ever raised a Series A, every enterprise software company that has retooled its marketing to include the words “machine learning”, “AI”, and “large language model”, and every consulting firm that has launched an AI practice, somehow, despite all of those, we still struggle to reach $50 billion in verifiable top-line revenue attributable to this generative AI wave.

Let us do the arithmetic step by step. At least $730 billion in annualized CapEx. Roughly $25-$30 billion in attributable revenue at the hyperscaler level, that is a CapEx to revenue ratio of at least 28:1. For every dollar of revenue the industry can actually point to, it is spending twenty-eight dollars on infrastructure. Even if we assume those investments have a five-year useful life and a seven-year depreciation schedule, both extremely aggressive assumptions for technologies that become obsolete with each new iteration of GPU, the ROIC is abysmal.

Compared this to the cloud computing buildout that the same companies undertook between 2012-2020. During that period, Amazon, Microsoft, and Google spent approximately $400 billion in cumulative CapEx on cloud infrastructure. By 2020, AWS, Azure, and GCP were generating over $100 billion in combined annual revenue. The CapEx to revenue ratio at the peak of that buildout cycle was roughly 4:1. And the ROI was never in doubt, enterprises were visibly migrating and actively embracing workloads, CFOs could point to real cost savings, and the adoption curve was measurable in quarterly disclosures from companies like Snowflake, Salesforce, and Adobe.

The AI buildout has none of this. The most commonly cited “demand signal” is the fact that end users can now ask ChatGPT or Claude…etc to draft a comprehensive and detailed email, use GitHub Copilot to autocomplete a demo prototype, or generate a “good-enough” video of a cat in Renaissance clothing. These are surely impressive technological demonstrations, and proof of a working concept, but they are not, in aggregate, a $730 billion a year addressable market.

The gap between narrative and reality has become so vast that it has reminded us of the observation attributed to the late John Kenneth Galbraith: “The only function of economic forecasting is to make astrology look respectable”.

III. The Circular Financing Problem: When Demand Is Manufactured by the Supplier

We have spent considerable time tracing the cash flow that underpin the AI industry’s demand narrative. What we and many other research analysts found is deeply troubling.

Consider the following stylised fact pattern:

Microsoft invests $13 billion in OpenAI, with the agreement that OpenAI will use Microsoft’s Azure cloud infrastructure exclusively.
OpenAI, now flush with plenty of cash, enters into a multi-year, multi-billion dollar computing agreement with Microsoft Azure.
Microsoft reports that Azure AI revenue is growing at 62% YoY and that the division is “capacity-constrained”.
Wall Street analysts, seeing revenue growth and hearing the previous phrase, raise their price targets and recommend that investors buy more Microsoft stock.
Microsoft, its stock price raised and buoyed, issues debt and allocates more capital to AI data center constructions, a portion of which will be used to serve growing computer needs of … OpenAI.

Are we seeing the issue here? In this simple illustrative and real life case, the same capital is being counted twice. Once as an investment from Microsoft’s balance sheet into OpenAI, and once as revenue from OpenAI flowing back to Microsoft’s income statement. This is not organic enterprise demand, this is a financial engineering loop masquerading as a technological revolution.

OpenAI has, by our count, announced over $1.4 trillion in aggregate computing commitments and strategic partnerships since 2023. A careful reading of the fine print reveals that much of this “demand” is through financing, specifically equity investments, convertible notes, and preferred stock purchases, that happens to be structured in a way that appears as cloud revenue on the hyperscalers’ books.

We are not accusing anyone of fraud though we note with interest that the SEC has recently expanded its enforcement division’s focus on AI-related disclosure issues. We are simply pointing out that when we trace the cash, a significant portion of the described “insatiable demand” for AI compute is actually recycling the same pool of hyperscaler capital through a series of related-party transactions.

The emperor’s new clothes are being sewn by the emperor’s own tailor.

IV. The Adoption Mirage: Where Is the Enterprise?

Every quarter, we read the earnings transcripts. Every quarter, we hear the same refrain: “Enterprise adoption of AI is accelerating”. And every quarter, we look at the actual data and find remarkably little to no evidence which support this claim.

Let us consider the state of enterprise AI adoption in Q3’2026:

The “Pilot Purgatory” problem. A survey conducted by Gartner in late 2025 found that over 70% of enterprise AI projects had not progressed beyond the pilot stage. This is consistent with our own conversations with CIOs and technology officers at medium to large multinational corporations. “We are experimenting” sums up our conversations with executives, they are running concept-to-proof deployments, testing copilot integrations, in most cases, they are at the stage of evaluating use cases for AI. Out of all our discussions, we have not heard any executives stating they are deploying at scale. No one is replacing existing workflows, and they are not generating measurable cost saving or revenue uplift to justify the cost in dollars their cloud providers are asking them to commit.

The hallucination ceiling. The fundamental technical limitation of LLMs, their tendency to generate plausible sounding but factually dubious if not false outputs, has not been solved. It has been mitigated at the margins by techniques such as retrieval-augmented generation and iterations of find-tuning, but the core problem persists with no solution for the root causes. In almost all of the enterprise applications where accuracy matters, such as healthcare diagnosis, legal document review, financial reconciliation and modeling, compliance monitoring, engineering calculations, the cost of verifying AI output often exceeds the cost of doing the work manually. The technology is impressive, but it is not reliable enough for the enterprise workloads that can slightly justify its buildout costs.

The security and regulatory barrier. GDPR, HIPAA, CCPA, the EU AI Act, China’s algorithm governance regulations, and a growing patchwork of national and sub-national laws are creating a compliance nightmare for any enterprise that wants to run sensitive workloads on their-party AI infrastructure. The legal risk of exposing customer data to a model that cannot guarantee information deletion, cannot certify data provenance, and cannot provide auditable reasoning chains is one that most general counsel are unwilling to accept, let alone adopt. Until this changes and there is no clear path to it changing, the enterprise market for AI will remain a fraction of what the hyperscalers are projecting.

The pricing paradox. The cost of inference or the actual act of running an AI model to serve a user’s request has been falling dramatically as more efficient architecture emerges and the market becomes more competitive. This is, on the surface, a good thing for the general consumers. But it is a catastrophic thing for anyone who has a multi-billion-dollar bet on selling inference compute capacity at premium prices. When DeepSeek demonstrated that a well-optimised model could achieve comparable results to GPT-4 at a fraction of the costs, it sent a signal that rippled through the entire industry (this is happening again with Kimi K3 recently): the marginal costs of AI will trend towards zero. If this is true, and we believe it is, the hyperscalers are building trillion dollar worth of infrastructure to sell a commodity that will eventually flatline toward 0. That is not an investment, that is a subsidy for the end consumer, paid for and by shareholders.

V. The Diminishing Returns Problem: When the Customer Does the Math

There is a version of this story in which the frontier AI labs are the greatest businesses of the century. We want to steelman it properly, because it deserves to be steelmanned and because the market is already pricing it in.

OpenAI’s most recent funding round valued the company above $1 trillion, a figure that if it were GDP metric, would rank it among the twenty largest economies on earth. Anthropic’s latest raise came in at nearly a trillion, $965 or so billion. xAI, a company that did not exist four years ago, is valued at approximately $250 billion during its SpaceX trading debut. The combined market value of the three leading frontier labs now exceeds $2 trillion against combined revenue we estimate at below $30 billion, a multiple north of fifty times revenue, before any adjustment for the fact that a meaningful share of that revenue is purchased, directly or indirectly, with hyperscaler capital. To compound into those valuations, the labs must grow from $30 billion in revenue today to several hundred billion within five years. That is not a forecast but an assumption, the assumption that AI expenditure becomes one of the largest line items in global corporate IT budgets, larger than enterprise security, larger than data infrastructure, larger than the payroll of entire engineering organisations.

And here we concede the point, that the pace of development is real. The frontier models shipping in 2026 are meaningfully better than those of 2025: better at reasoning, better at coding, better at the long-horizon tasks that were hopeless eighteen months ago. If the technology is as transformative as its proponents claim, the demand curve could be far steeper than the current run-rate suggests. The bulls may well be right that the labs can hit their numbers. Our concern is not whether the labs can hit their numbers. It is what happens to the customer while they do.

The ROI curve inverts. Every enterprise purchase is, at bottom, an ROI decision, and the question every customer eventually asks is not “is this model better?” but “is this model better enough to justify the price?” Here the recent record is not encouraging. By our estimates, the last three generations of frontier models have delivered single-digit percentage improvements on the enterprise benchmarks that really matter: accuracy on complex workflows, reliability on multi-step tasks, while API prices per token have roughly doubled or tripled with each release. The customer is being asked to pay exponentially more for arithmetic progress. At some point on that curve, nearer than the industry admits, the marginal token costs more than the marginal value it creates. That is the definition of diminishing marginal returns. It arrives at a different point for every customer, but it arrives for all of them eventually.

The $20 arbitrage. Consider the current state of frontier pricing, which has become genuinely absurd. One lab charges enterprise customers premium API rates for its flagship model, Fable 5, at prices north of $100 per million output tokens. A modest enterprise workload consuming a hundred million output tokens per month therefore carries a six-figure annual price tag before integration, before fine-tuning, before the human verification layer we will come to in a moment. The competing frontier model, Opus 5, benchmarks comparably on every evaluation that matters, and it is available to anyone with a credit card for $20 a month. Twenty dollars. The annual API bill for that single modest workload exceeds five hundred years of consumer subscriptions.

Why would any rational CFO sign that contract? The labs have an answer, and it is not a bad one: enterprise support, service-level agreements, data controls, rate limits, indemnification. All of that is real, and all of it is worth a premium. But it is not worth a five-hundred-year premium. The moment a competitor offers comparable frontier capability at a saner price and open-source models are already circling at effectively zero, the premium collapses. The labs are simultaneously trying to maximise revenue per customer and defending prices that the most casual consumer comparison renders indefensible. You cannot charge monopoly rents for a product your competitor sells for the price of a streaming subscription.

The oversight tax. There is a line item in every enterprise AI deployment that appears in none of the marketing materials: the human verification layer. A bank does not let a model approve a loan without a human signing off. A hospital does not let a model discharge a patient without a physician reviewing the reasoning. An engineering team does not ship model with generated code without a senior developer reading the diff, because the hallucination ceiling documented earlier in this essay has not been solved, only pushed higher. That layer is not a rounding error. In the deployments we have examined, it consumes as many human hours as the work being automated, and it is frequently the most expensive component of the entire system.

Add the verification layer to the token bill, and the fully loaded cost of AI-assisted work routinely exceeds the fully loaded cost of the human doing the work directly. We have spoken to engineering department members who confess, off the record, that their “AI-augmented” teams now cost more per delivered feature than the pre-AI teams did and that no one says so publicly, because the capex was approved at the board level and the incentives run the other way. The emperor is not naked. He is wearing an extremely expensive suit he cannot return.

The return of the human. Which brings us to a prediction we make with unusual confidence. At some point on the current pricing trajectory, closer than the industry believes, the CFO’s spreadsheet will recommend hiring ten engineers rather than renewing the API contract. The marginal cost of a human is stable; the marginal cost of frontier intelligence, at current pricing, is rising. When the curves cross, the market for AI labour inverts. After a decade of headlines predicting the obsolescence of the software engineer, the equilibrium is a return to human engineers, not because the models are bad, but because the models, once you pay for frontier access and the humans required to verify their output, are more expensive than the engineers they were supposed to replace. The humans return as the cheaper option. That was not the plan.

And this is the hidden flaw in the bull case. The revenue the market is capitalising must be extracted from customers who are, quarter by quarter, doing this arithmetic and finding the answer negative. When enough of them defect to cheaper models, to open-source alternatives, or back to human beings, the growth that justifies the valuations evaporates. The labs may hit their numbers for a year, or two, or three. But the customer-side economics break first, and when they break, the revenue breaks with them. The bull case and the bear case converge on the same destination: a revenue cliff, approached from opposite directions.

V. The Self-Fulfilling Prophecy: “Spend Enough and Demand Will Follow”

The most honest articulation of the AI investment thesis we have encountered came not from a CEO or a sell-side analyst, but from a venture capitalist at a Chinese VC conference that we attended earlier this year. Off the record, he said: “We’re spending so much that we’re going to force demand into existence. There is really no going back to gradualization”.

This is the self-fulfilling prophecy in its purest form. The logic proceeds as follows:

The hyperscaler collectively spends hundreds of billion on AI infrastructure.
The suppliers of that infrastructure: NVDIA, AMD, Broadcom, TSMC, the data center REITs, the electrical grid operators, the construction firms will see an enormous boom in their own revenue, valuation and growth projections.
The stock prices of these suppliers rise, enriching the very investors whose capital fund the buildout.
Because so much capital is now sunk, the hyperscalers must find something or rather anything, to put on top of this infrastructure that generates sufficient traction to justify the spending, FYI we are still not talking about revenue here.
Startups and enterprises, sensing the availability of cheap compute and the strategic imperative to “do something with AI”, begin building applications.
Some of these applications find genuine products to market fit, but most will not.
The hyperscaler uses this as an example of proof that their bet was in fact correct and is “paying off”.
The cycle continues, no one can afford to admit that the first several hundreds of billions (potentially even a trillion) was speculative.

There is actually a word for this way of thinking, we have seen this pattern over and over again in financial history. It is called pump and dump, and it has been around since the days of the South Sea Company. The difference is that this time, the promoters are not boiler-room operators in Florida, they are the largest, most advanced, and perhaps the most respected companies in the world, and they are using their own shareholders’ capital to manufacture the demand that justifies their own proof-of-concept thesis.

The South Sea Company, if we recall, was formed in 1711 to trade with Spanish colonies in South America. It had essentially no revenue, no competitive advantages, and no viable business model. But the British government granted it a monopoly, and the stock rose more than 800% in a single year as investors convinced themselves that the promise of South America trade would materialize if only enough capital were committed. Isaac Newton, who initially profited from the mania, later lost the equivalent of today’s 40 million pounds when the bubble burst. His reported comment: “ I can calculate the motion of heavenly bodies, but not the madness of people”.

We want to emphasize that we do not believe AI is a fraud in the sense that the South Sea Company was a fraud. The underlying technology is real. It has genuine, transformative potential in certain applications. But the gap between the technology’s current capabilities and prices the market is demanding for exposure is as wide as any we have observed in most of our short careers. And the capital that is being deployed to close this gap, the annual CapEx, is being spent not because the ROI is calculable, but because the fear of being wrong is greater than the discipline of being right.

If we think about it, hyperscalers really are treating shareholders with no basic respect (in a lack of better term, they are treating investors and shareholders like idiots), shareholders are basically using their own money to test out a concept forced upon them, even if AI makes revenue, the amount of spending likely means shareholders will be underwater for the next 20 years.

VI. The Fraud Landscape: Why Trust Is the Unspoken Variable

We have written at length elsewhere about specific instances of fraud and misconduct within the AI ecosystem previously. But they bear repeating here, because they are not anomalies, they are a common symptom of a structural incentive problem. We decided to list couple from our numerous internal memo over the past 2 year:

Super Micro Computer: One of the largest beneficiaries of the AI infrastructure buildout, with a market capitalization that briefly exceeded $70 billion, has been under investigation by the DoJ and the SEC for accounting manipulation, undisclosed related-party transactions with entities controlled by the CEO’s brothers, and suspected violations of export controls. Its auditor, Ernst & Young (EY), resigned in October 2024, stating that it was “unwilling to be associated” with management’s financial statements and citing concerns about “integrity and ethics” (quite ironic coming from EY, regarding their frequent “integrity and ethical” misconduct). The stock crushed 33% in a single day, erasing $10 billion in market value. A replacement auditor, BDO, subsequently issued an “adverse opinion” on the company’s internal controls. This is not a minor accounting disagreement, this is a company at the center of the AI hardware ecosystem that has been caught, for the second time, manipulating its financial statements.
Ostin Technologies Group: A small Chinese manufacturer of LCD components with $38 million in annual revenue and a negative profit margin, its stock surged over 1,175% in two months after a coordinated pump and dump scheme using AI-generated deepfake video of Elon Musk and Mark Zuckerberg, stolen FINRA-registered advisor identities, and WhatsApp groups to manufacture the appearance of investor demand. When the scheme collapsed on June 26, 2025, the stock fell from $9.40 to $0.55 in a single session, a 94% decline, which wiped out close to a billion dollars in market capitalization. The co-CEO has been indicated by the DoJ. The company’s shares have been suspended ever since by NASDAQ.
Nate Inc. A startup that raised $42 million from sophisticated VC investors (which included Forerunner Ventures, Canaan Partners and Coatue Management) by claiming its app could use AI to complete purchases from any online store with a single tap. The reality, according to federal prosecutors: hundreds of human workers in the Philippines and Romania were manually completing every purchase behind the scenes. Not one transaction was automated. The CEO, Albert Saniger, was charged with securities fraud and write fraud by the DoJ and the SEC in April 2025. Investor lost everything.

AI washing across the industry. A February 2025 survey by MMC Ventures found that 40% of fintech startups branding themselves as “AI-first” had zero machine learning code in production. A quarter were simply piping third-party APIs through a new user interface. The SEC and the FTC have launched dedicated enforcement operations targeting false AI claims. The FTC brought at least a dozen such cases in 2025 alone. And yet the term “AI-powered” continues to command a 2-3x valuation premium in private markets.

We cite these cases not to suggest that every company in the AI ecosystem is fraudulent, in fact the majority of them are not, but to illustrate a structural transparency problem. The industry is characterised by:

Growth traction matrices that cannot be transferred into recurring and stable revenue: most agentic providers allow users to use limited functioned models for free, but struggle to turn those existing users into paying subscribers.
Revenue that is difficult to verify: when a cloud provider says its AI revenue is growing at 62%, how much of that is genuine third-party enterprise demand and how much is related-party recycling?
Costs that are deliberated opaque: no hyperscaler provides a clear breakdown of AI CapEx, maintenance CapEx and other miscellaneous CapEx. No one discloses utilization rates for their GPU clusters. No one explains what happens to the billions in hardware investment when the next iteration of more efficient chips renders the current generation obsolete.
Metrics that are carefully curated: “token volume”, “API calls”, and “model parameters” are not GAAP financial metrics. They are marketing numbers. They are chosen precisely because they cannot be easily audited or compared.
Insiders who are selling: we have tracked insider trading patterns at AI-adjacent companies and found elevated levels of insider selling that are inconsistent with the public narrative of boundless possibility and insider optimism/confidence.

A market that cannot be seen clearly is a market that cannot be priced correctly. And a market that cannot be priced correctly is a market that will eventually correct violently, as the gap between perception and reality becomes too wide to sustain.

VII. The Agency Problem: Who Actually Benefits From This Spending?

Let us ask an uncomfortable question that is rarely posed on earnings call: are the executives making these capital allocation decisions properly aligned with their shareholders”？

We have spent considerable time examining the compensation structures, insider trading patterns, and personal financial incentives of the CEOs, CFOs, and CIOs who are directing this massive spendings. What we have found makes us deeply uneasy.

Stock-based compensation rewards grand narratives, not capital discipline. The typical CEO of a Mag 7 company holds the vast majority of their personal wealth in company stock and stock options. Their compensation is structured around total shareholder return relative to a peer group, the very same peer group with which they are now locked in an AI CapEx arms race. This creates a powerful incentive to spend aggressively not because the ROI justifies it, but because not spending would cause the stock to underperform the peer group.

Consider mathematics. A CEO who announces a $100 billion CapEx programme and sees the stock rise 10% on the narrative earns multiples of their annual salary through option appreciation. A CEO who announces capital discipline, sits out the AI arms race, and sees the stock stagnant or even fall 10% as investors worry about being “left behind” loses personal wealth. The asymmetry is stark. The system is designed to reward boldness, not providence. And when CEO tenure at the largest tech companies average less than seven years, the incentive to pursue short-term narrative lift over long-term value creation is overwhelming.

Insider selling is flashing red. We have tracked insider trading patterns, as mentioned above, found a pattern that deserves more attention. At NVIDIA, insiders sold over $1.8 billion in stock during 2025, the seller includes CEO Jensen Huang, who executed a series of pre-planned 10b5-1 sales that, cumulatively, reduced his personal exposure by hundreds of millions of dollars. At Meta, insider sales accelerated throughout 2025 into 2026, even as management was publicly projecting boundless optimism about AI monetisation. At Amazon, founder Jeff Bezos sold over $6 billion in stock during the calendar year of 2025.

We are not suggesting any of these transactions are illegal. The vast majority are conducted through pre-arranged trading plans precisely because the executives are so awash in stock-based compensation that any portfolio diversification requires regular sales. But the pattern is worth taking notes: the people with the best information about whether massive annual AI CapEx will generate returns are, in aggregate, reducing their personal exposure to the very equities they are asking shareholders to invest more in.

The golden handcuff paradox. The most talented AI researchers, the people who would know if the technology is approaching fundamental limitations, are earning eight figure compensation packages to stay at the hyperscalers. They have no incentive to publicly question the narrative (why would they bite the hand that feeds them?). The few who have raised concerns internally about the scaling laws plateauing, about the diminishing returns of large models, or about the reliability ceiling of current architectures, have either been promoted into non-technical management roles or have quietly departed for well-funded startups where the compensation is even higher. The information asymmetry between what the engineers know and what they market prices is vast, and every structural incentive in the industry pushes toward maintaining the illusion rather than puncturing it.

The debt markets are enabling the behaviour: we observe with concern that the hyperscalers are increasingly funding their AI CapEx through debt rather than operating cash flow. Alphabet quadrupled its long-term debt to $46.5 billion in 2025. Amazon has filed with the SEC to potentially access the equity and debt markets. Moody’s has flagged the 94 % CapEx to cash flow ratio across the sector as credit turns negative. The bond market, in its current yield-starved state, is happily absorbing this issuance at negligible spreads. But credit cycles turn, when they do, the companies that used cheap debt to fund speculative infrastructure will find themselves trapped between maturing obligations and disappointing AI revenue. The debt markets are enabling an overinvestment cycle that they will eventually be forced to price correctly.

We believe the agency problem is the most underappreciated risk in the AI infrastructure narrative. The people making the decisions are not bearing the full cost of being wrong. The shareholders who fund the spending, the employees who contribute their labour, and the communities that host the data centers are the ones who will absorb the losses. The executives and founders, meanwhile, have already diversified.

VIII. The Open-Source Commoditization Threat: Where Is the Moat?

The central premise of the hyperscaler AI investment thesis is that the companies building the largest and most capable AI model will capture a disproportionate share of the economic value generated by those models. This is the “winner-take-most” hypothesis, and it is the intellectual foundation upon which $730 billion in annual CapEx rests.

We believe this premise is flawed. The evidence increasingly suggests that AI models are not a winner-take-most market but a rapidly commoditising one.

Open-source models are closing the gap. Neta’s Llama family, Mistral’s open-weight models, and most major Chinese AI companies’ breakthroughs in training efficiency have demonstrated that frontier-competitive AI can be built, trained, and distributed by organisations spending a fraction of what the hyperscalers are spending. DeepSeek-V3, released in late 2024, achieved performance comparable to most Western models on several benchmarks costing an estimated $6 million to train; this is less than 0.01% of what OpenAI and Microsoft have collectively spent on frontier model development up to that point. The gap between proprietary and open models is measured in months, not years. And it is shrinking exponentially, demonstrated by the release of Kimi K3 of the Moonshot AI, its benchmarks and pricing are potentially able to outcompete Anthropics’ Fable 5 in the near future.

Commoditization is accelerating, not decelerating. Every subsequent model release in 2025 and 2026 has reinforced our understanding. The incremental gain from scaling to larger models, the famed “scaling laws” that underpinned the industry’s thesis, is diminishing. Smaller, more efficient architectures are achieving comparable results through better data curation, more efficient training techniques, and innovative inference-time strategies. The cost of training a frontier-quality model has fallen by an order of magnitude in the past 18 months. The cost of inference has fallen even faster.

If the model itself is not the moat, what is? The hyperscalers would answer: distribution, enterprise relationships, and the integrated workflow stacks. We have examined this argument and find it unconvincing. If a mid-sized enterprise can download an open-source model, run it on commodity hardware or third-party cloud infrastructure, and achieve results comparable to GPT-5 for a fraction of the cost, the hyperscaler’s proprietary advantage completely erodes. The “integrated workflow stack” argument, that customers will pay a premium for the convenience of having model, compute, and application in one place, only holds if the premium is small relative to the value delivered. At current pricing levels, the premium is not small. It is in fact enormous. And open-source alternatives are making it more visible by each quarter.

The hyperscalers are caught in a contradiction they cannot shake off. They are spending so much to train ever-larger proprietary models while simultaneously investing in open-source ecosystems. Meta releases Llama as open-weight. Google has open-sourced Gemma. Microsoft has invested in open-source AI infrastructure. This is not altruism, it is a product hedge. Deep down, they know (the decision makers) that the proprietary moat will not hold, and they are positioning themselves for a world where models are cheap (if not free) and abundant (if not everywhere). The problem, from an investment perspective, is that a world of cheap and abundant models does not support the current valuation of companies within the AI industry, and it further contradicts massive CapEx for building infrastructure to serve sub-par models at premium prices.

IX. The Geopolitical Powder Keg: One Island, One Company, One Point of Failure

The entire global technologies and AI infrastructure buildout rests on the assumptions that the global supply chains for advanced semiconductors will function uninterrupted for the foreseeable future. This assumption deserves far more scrutiny than it has received.

Taiwan Semiconductor Manufacturing Company Limited (Chinese: 台灣積體電路製造股份有限公司 ; Taiwan Semiconductor or TSMC) is the single point of failure. Every advanced AI chip, from every NVIDIA GPU to every AMD accelerators, even Google’s TPU…etc is manufactured by a single company in a single country. The advanced packaging technology that enables these chips to function, CoWoS, is even more concentrated, with TSMC controlling the vast majority of the global capacity. A disruption at TSMC, whether from a geopolitical conflict in the Taiwan Strait, a natural disaster, a pandemic-related shutdown, or a technical failure at its fabrication facilities in Hsinchu and Tainan would halt AI chip production for months, possibly longer. There is no space capacity, at least for now, no sustainable Plan B.

The friend-shoring narrative is aspirational, not operational: the US government and hyperscalers have announced ambitious plans to build semiconductor capacity across Arizona, Texas, Ohio, Japan, and Germany, but assuming these facilities come online on schedule is heroic given the industry’s construction record, and even then they will not meaningfully reduce dependence on TSMC for at least five to seven years, likely closer to a decade, while producing older-generation nodes rather than the bleeding-edge processes AI chips require, meaning supply-chain concentration risk is compounded rather than mitigated by the sheer demand volume placed on that single gear. Employee interviews confirm TSMC’s demanding culture, including 12-hour shifts, high burnout, and elevated turnover among local teams, and its US expansion has indeed been slower and costlier, with management conceding that building fabs in Arizona takes at least twice as long as in Taiwan and that construction, permitting, workforce, and calibration challenges pushed Fab 2 volume production to H2 2027.

Yet the inference of wide operational inefficiency is unsupported: yield at the Phoenix fab is at parity or better than comparable Taiwan lines (reported 92% on 4nm, several points above Tainan, with C.C. Wei stating yields are approaching Tainan levels), first US-made Blackwell-class wafers are already in production, and the gap is explained not by worker culture but by TSMC’s proven playbook of transferring process recipes and ramping new fabs with experienced Taiwanese engineers before handing over to local staff, a model further validated by the ahead-of-schedule Kumamoto ramp in Japan. The genuine Arizona handicap is therefore economic rather than technical: roughly 2x construction timelines, wafer costs estimated 35 to 50% above Taiwan, and a supply chain incomplete until domestic advanced packaging (Amkor’s Peoria facility) comes online, which is why TSMC keeps investing ($265B across up to 10 fabs) despite the premium; customers are paying for geographic diversification and AI-driven demand, not US cost parity. The bear case on friend-shoring rests on speed and cost, and it is strong; the bear case on TSMC’s people and technology has so far been falsified by the data, as the dependency the market fears persists through demand concentration and build-out delays, not through any degradation of TSMC’s operational competence outside Taiwan.

Export controls are double-edged. The U.S. government’s restrictions on exporting advanced AI chips to China have, in the short term, protected the hyperscalers’ domestic position. But they have also created powerful incentives for China to develop its own semiconductor ecosystem, and for Chinese technology companies to achieve AI breakthroughs using less advanced hardware. The emergence of the Chinese hyperscalers, which demonstrated frontier-competitive performance on restricted hardware, should be read as an early warning. The export controls have not permanently disadvantaged China. They have accelerated development of a parallel semiconductor ecosystem that will, given enough time, compete directly with western foundries.

The energy constraint is geopolitical, not just technical. AI data centers are projected to consume between 5 to 10% of total U.S. electricity generation by 2030, up from approximately 3% today. This demand is colliding with the requirement of coal-fired power plants, the intermittency of renewable energy sources, and decades of underinvestment in grid infrastructure. The hyperscalers are signing nuclear power purchase agreements: Google with Kairo Power, Microsoft with Constellation Energy, Amazon with X-energy, but these facilities will not come online for many more years, if we assume they come online at all. In the interim, data centre construction is being slowed in multiple jurisdictions by utility companies and local governments that cannot guarantee the required power supply. The infrastructure buildout is hitting the physical limits of the electrical grid, and those limits are not easily solved by throwing money at them. Electricity takes time, regulation, and physical construction. All three are in short supply.

X. The Energy Mirage: The Grid Cannot Support the Narrative

Let us dwell on this point for a moment, because it is the least discussed and most physically constraining of all the risks facing the AI industry.

The numbers are staggering. A single AI data center campus, like the ones Meta is building out in El Paso, Texas, is projected to consume 1 gigawatt of power when fully operational. Microsoft’s Hyperion campus in Louisiana is designed to deliver 5 gigawatts, enough to power approximately 3.5 million homes. The combined power draw of all AI data centers under construction or planned in the U.S. as of mid-2026 is estimated at over 40 gigawatts. The entire U.S. grid has approximately 1,200 gigawatts of installed capacity. The AI industry is effectively demanding a 3-4% increase in total U.S. electricity generation capacity within a five-year window and this in a country that has not built a new large-scale nuclear reactor in decades, where natural gas faces growing environmental opposition, and where renewable projects face years of interconnection queues.

The carbon contradiction is irreconcilable. Every hyperscaler has made ambitious net-zero commitments. Microsoft has pledged to be carbon-negative by 2030. Google has committed to 24/7 carbon-free energy by 2030. Amazon has committed to net-zero carbon by 2040. Yet the same company is building data centers that will each consume many gigawatts of electricity, requiring either a natural gas backup generation, new nuclear construction that will not materialize in the relevant timeframe, or an expansion of renewable energy that will be diverted from grid decarbonization. The carbon footprint of the AI buildout is already larger than the entire aviation industry’s emissions in several major economies, and it is growing exponentially. The net-zero commitments and the AI CapEx commitments are in direct contradiction. One of them will be abandoned, and we suspect it will be the climate commitments.

Local communities are pushing back. Data center construction is facing growing opposition from local residents, environmental groups, and municipal governments concerned about water consumption, noise pollution, diesel generator emissions, and the strain on local power grids. Northern Virginia, the world’s largest data center market, has seen local zoning battles over AI facility construction. The Netherlands imposed a moratorium on new data centers. Singapore lifted its moratorium but imposed strict energy efficiency requirements, Ireland’s grid operators have warned that data center demand could exceed supply. The political tailwinds that the AI industry has enjoyed are not guaranteed to persist. As the externalities, from higher electricity prices, water scarcity, to carbon emissions become more visible, the regulatory response will have to tighten. And tighter regulation means slower buildout, higher costs, and reduced returns.

XI. The Regulation Tsunami: The Costs Nobody Has Priced

Let us dwell on this point for a moment, because it is the least discussed and most physically constraining of all the risks facing the AI industry.

The AI industry is currently enjoying a regulatory vacuum. The technology has advanced far faster than the laws and regulations that govern it. This vacuum will not persist. And when regulation arrives, as it is already arriving in Europe and China, the costs will be material.

The EU AI Act is already passed as a set of law. The European Artificial Intelligence Act, which entered into force in stages throughout 2025 and 2026, imposes sweeping requirements on any organization that develops or deploys AI systems within the EU market. These include mandatory risk classification, conformity assessments, transparency obligations, human oversight requirements, and record-keeping standards. Non-compliance can result in fines of up to 70% of global annual turnover, a penalty structure that applies to any company, European or otherwise, that serves EU customers. The compliance costs for a large hyperscaler operating dozens of foundation models across hundreds of use cases will run into the billions of dollars annually. This cost has not been factored into the AI revenue projections that justify the CapEx.

The U.S. regulatory patchwork is fragmenting. There is no comprehensive federal AI law in the U.S. Instead, a growing number of states are passing their own AI legislation, such as covering algorithmic discrimination, deepfake transparency, automated decision making, and AI in hiring, housing and healthcare. By mid-2026, more than 25 states have enacted or are actively considering AI-related legislation. These laws are not harmonised. A company deploying an AI system across the U.S. must comply with a growing and inconsistent patchwork of state-level requirements that differ on definitions, obligations, and enforcement mechanisms. The legal complexity and compliance cost of operating at scale in this environment are significant and will only increase.

China’s algorithmic governance is tightening. The PRC has implemented one of the world’s most comprehensive framework for regulating AI, including the Algorithmic Recommendation Provisions, the Deep Synthesis Provisions, and the Generative AI Measures. Foreign companies operating in or serving the Chinese market, or using AI systems developed in China must navigate a regulatory environment that requires algorithm registration, security assessment, content control compliance, and government access to model information. The regulatory risk for any hyperscaler with a cross-border AI business is substantial and insufficiently priced.

Intellectual property litigation is a growing overhang. The copyright, patent, and trade secret challenges to foundation model training are multiplying. Authors, visual artists, software developers, news publishers, and stock photo agencies have filed class-action lawsuits against OpenAI, Microsoft, Meta, Google, Stability AI, and others, alleging that their models were trained on copyrighted works without permission or compensation. The legal theories are novel and untested. The potential damages are enormous: statutory damages for willful copyright infringement can reach $150,000 per work, and the training datasets used by major foundation models contain billions of copyrighted works. Even a fraction of successful claims could result in liability that dwarfs the current revenue of the AI industry. No hyperscaler has set aside meaningful legal reserves for this risk. They are betting that fair use will carry the day. That bet is far from certain.

XII. The Magnificent Seven: A Valuation Autopsy

Let us now turn to the stock themselves, because valuation matter so much in the current cycle. The Magnificent seven or Mag 7: Apple, Microsoft, Alphabet, Amazon, Meta, NVIDIA, and Tesla account for over 35% of the entire S&P 500 by market capitalization. At the peak of the dot.com bubble, the five largest technology stocks accounted for roughly 18%. The concentration of market power and investor capital in a handful of AI-narrative stocks is historically unprecedented.

The most concerning metric is the relationship between CapEx and FCF. Consider what has happened to FCF generation at these companies as they have scaled their AI investments:

Alphabet’s FCF is projected to decline from $73.3 billion in 2025 to approximately $8.2 billion in 2026, a 90% collapse.
Meta’s FCF is also projected to fall by approximately 90%. Barclay analysts have modelled negative FCF for Meta through 2028.
Amazon’s FCF is expected to turn negative, with BoA estimating a deficit of $17 top $28 billion in 2026.
Microsoft’s FCF is more resilient, declining by an estimated 28%, but Microsoft’s absolute spending is also hardest to pin down, with estimates ranging from $105 billion to $190 billion depending on fiscal year assumptions.

What happens when the largest companies in the world stop generating FCF? They have to issue debt, they also need to halt buybacks and any dividends. They cannibalize other core business units, Meta is already doing all of these things simultaneously.

The equity market, in our view, has not fully priced the risk that these capital commitments will permanently impair the compounding power of these businesses. A company that once returned $50 billion annually to shareholders through buybacks and now must redirect that capital into GPU clusters is a fundamentally different investment (we would argue any company that’s doing this extensively is misallocating shareholder’s investments). The multiple that investors assign to a cash flow generating machine should not be the same multiple they assign to a capital intensive infrastructure project with uncertain returns.

XIII. The NVIDIA Problem: One Company, 90% of the Hardware Market

No discussion of AI CapEx would be complete without addressing the elephant in the data center: NVIDIA.

NVIDIA currently commands an estimated 90% market share in AI training and inference GPUs. It enjoys gross margins of approximately 74.5%. Its market capitalisation, as of this writing, is comfortably above $3 trillion, placing it among the three most valuable companies in the world.

NVIDIA’s revenue growth over the past three years has been extraordinary. But the question that keeps us up at night is not whether NVIDIA has had a good run. It is whether the hyperscalers who are NVIDIA’s largest customers, the same companies investing hundreds of billion in AI infrastructure are building the very capacity that will, in time, reduce their dependence on NVIDIA’s hardware.

Because that is exactly what they are doing. Amazon is designing custom Trainium and Inferentia chips. Google has its Tensor Processing Units. Microsoft has co-developed the Maia chip. Meta is working on custom silicon. So is the company under Elon Musk. Every hyperscaler is investing billions in reducing their reliance on NVIDIA’s proprietary ecosystem and its pricing power. If even a fraction of these efforts succeed, NVIDIA’s market share will erode violently, its margin will compress, and its revenue growth will decelerate. We only listed a couple U.S. companies developing their own silicon, think in geographies, NVIDIA has a large presence in China (even after both countries restrictions), a large portion of revenue comes from China, what if China becomes chip independent?

This is, from an industrial logic perspective, perfectly rational. No customer willingly pays 74.6% gross margins forever if they are able to vertically integrate. But for investors who are pricing NVIDIA as if its current growth trajectory will persist indefinitely, the risk is portfolio existential.

There is another layer to this that receives insufficient attention: concentration risk in the supply chain. NVIDIA’s GPUs are manufactured exclusively by TSMC (as discussed above), the Taiwanese semiconductor foundry. The advanced packaging technology that NVIDIA requires, CoWoS, is capacity-constrained and available from a very small number of suppliers. A single geopolitical event in the Taiwan Strait, a natural disaster affecting TSMC’s facilities in Hsinchu, or a capacity allocation decision by TSMC’s management could halt production for months, if not forever (in the case of China launching an invasion on Taiwan). The entire AI infrastructure buildout is resting on a supply chain that runs through one island, one company, and one set of manufacturing processes. That is not prudence, this is a spine-tingling lack of diversification.

XIV. The Alternative View: What If We Are Wrong?

A responsible portfolio manager must always consider the possibility that their thesis is incorrect. What if the AI optimists are right? What if these investments do generate trillions of dollars in economic value, and the companies that spent most on infrastructure are the ones that capture the most of that value?

This is entirely possible. We concede this. The internet buildout of the late 1990s was accompanied by massive overinvestment, remember the hundreds of billions spent on fiber-optic cable that briefly became known as the “dark fiber”? But the companies that survived and scaled ultimately generated enormous returns. Amazon was a dot.com era miracle story that looked ridiculous for years and then became one of the best investments in financial history.

We believe there are three reasons why the AI analogue to the internet buildout is flawed:

The internet had immediate, measurable consumer demand. People are signing up for AOL and CompuServe by the million. E-commerce, while initially a tiny fraction of retail, was demonstrating cheaper and more convenient than all the alternatives and traditionals. AI, by contrast, is solving a problem that most consumers and enterprises did not know they had. The demand is artificially manufactured by the very companies that are spending hundreds of billions on infrastructure. There is no organic groundswell comparable to “I want to email my friends” in 1994.
The internet’s infrastructure costs were shared and distributed. The fiber-optic buildout was funded by hundreds of telecommunication companies, a significant portion of those eventually went bankrupt. The surviving infrastructure was acquired for pennies on the dollar by the eventual survivors, or the “winners”. In the current AI buildout, four companies are funding hundreds of billions from their own balance sheets. There is no bankruptcy auction that will allow a more disciplined competitor to acquire these assets cheaply. The hyperscalers are eating the full costs of their own excesses.
The internet era’s excesses was followed by a brutal cleansing. The Nasdaq fell over 78% from its peak in March 2000 to its trough in October 2002. Trillions of dollars in market capitalization have essentially evaporated. Thousands of companies failed, and tens of thousands were laid off. The executives who presided over the overinvestments, such as the CEOs of WorldCom, Global Crossing, Enron all went to prison. In the current cycle, we see no mechanism for a similar cleansing. The four hyperscalers are too large, too systematically important, and too politically connected to be allowed to fail, they are “too big too fail”. The risk, therefore, is not the excess is purged through bankruptcy, but that it is socialized that all those CapEx spending is simply written down over time, imperceptibly, through years of disappointing returns that never quite trigger a crisis but permanently impair the compounding of the entire U.S. technology equity values. This is, in so many ways, the worst outcome for all U.S. facing equity investors. A crash is painful but finite (can even argue it is timed, controlled and managed), a slow decay of returns is interminable, the slow death of the UK equity market can be a cautionary warning for this scenario.

XV. Our Portfolio: The Intellectual Case for Zero Exposure

We are sometimes asked, usually by our LPs and peers, whether we are simply unskilled at picking good AI companies to invest in. The answer to us is a definitive NO. We are afraid of what happens when the last order bell rings.

Our mandate is not to capture every thematic wave. It is to compound capital over a multi-year, and increasingly multi-decade horizon with a disciplined approach to risk management and a healthy amount of skepticism of census narratives. We have examined the AI industry, we followed it for many years, its financial statements, its supply chains, its revenue attribution, its adoption curves, its regulatory environment, its fraud pattern, and we have concluded that the risk-reward calculus, unfortunately, is biased towards the house and not us the players.

This does not mean we refuse to invest in technologies, we wrote recommendation pieces on Meta previously, we even had a large position in Meta all the way up to last quarter (which we seem to mispriced Meta’s near-term performance). We still hold positions in companies that use software and automation to generate measurable, auditable returns in industries where the unit economics are transparent and adoption is real. We are interested in owning businesses in semiconductor capital equipment that do not depend on NVIDIA’s specific product cycle. We recommended enterprise software companies, and even enterprise AI companies whose customers can demonstrate a quantifiable return on their investment through the deployment of models. We own a small amount of infrastructure assets, REITs, energy, and logistics that will benefit from the buildout regardless of whether the AI thesis ultimately proves correct.

But we do not own NVIDIA. We do not own Microsoft, Alphabet, Amazon and now Meta in sizes that would expose us to their AI CapEx appetite. We do not own and will not participate in any SPACs, the pre-revenue startups, or special-purpose acquisition companies that have rebranded themselves as “AI platforms”. We do not own the stocks whose valuation depends on continuation of a narrative that, in our view, has disconnected from underlying economics.

We recognise that this positioning carries its own risks. We will underperform in the near term if the AI narrative continues to inflate. We accept this. The discipline of a family office is capital management for generations, is that we can afford to be early, but we cannot afford to be wrong.

And we believe, deeply, after years of observation, that the consensus view on AI CapEx, AI adoption, and AI revenue is wrong, from the very beginning. The spending is too high, the revenue is too low. The transparency is non-existent. The fraud is too prevalent. The self-fulfilling prophecy that enough spending must eventually create demand has never worked in the history of financial markets. If we are students of history, we should expect it will not work this time either.

The greatest misallocation of capital in human history is unfolding in real time. We choose to watch from the sidelines, not because we are afraid of technology, but because we are afraid of the price.

XVI. Epilogue: A Letter to Future Selves

We will conclude with a note we intend to read in a couple years time, when the AI narrative has either validated itself or collapsed under the weight of its own contradictions.

To us in 2030, from 2026:

If you are reading this and the AI industry has generated the returns its proponents promised, we ask only that you hold us accountable for our caution. Write a new edition of this diary explaining where we were wrong, and distribute it publically. We will take our medicine.

But if you are reading this in the aftermath of the correction we suspect is coming, if the CapEx has been written down, if the frauds have been exposed, if the revenue has failed to materialise and the valuation have compressed, we asked you to remember why we sat this one out.

We sat it out because we have seen this before. We have seen capital chased by more capital. We have seen a narrative that cannot be falsified. We have seen companies spending money they do not have on assets whose value they cannot measure. And we have seen that the end is always the same.

When the show is over, the curtain call comes; when the music stops, the margin calls will be put through. The paper profits evaporate, and the disciplined capital that waited that kept its powder dry and its wits about gets to buy assets at unbelievable prices. Prices that reflect reality and not hope.

That is what we are waiting for. Not the destruction of innovation, not the failure of a technology we believe has genuine potential, but the return of prices to a level where risk and reward are once again aligned.

Until then, we observe, we analyze and we write it down.

Disclaimer, Disclosure, Conflicts & Copyright Notice

This publication has been prepared solely for informational and educational purposes by Alpha Talon Investment Research (“Alpha Talon”). The views expressed herein represent the author’s independent analysis and opinions as of the date of writing and may change without notice. This material does not constitute investment advice, financial advice, legal advice, tax advice, or a recommendation to buy, sell, or hold any security, derivative, or financial instrument. Nothing contained in this document should be construed as an offer to sell or a solicitation to buy any securities.

Investing in securities involves significant risk, including the possible loss of principal. Equity investments may fluctuate in price, sometimes dramatically. Securities in the consumer discretionary and staples sector—such as those discussed herein—are subject to heightened levels of consumer, macroeconomics, competitive, and operational risk. Forward-looking statements, projections, price targets, valuation scenarios, and estimates included in this report are inherently speculative, based on numerous assumptions, and may differ materially from actual outcomes. Past performance is not indicative of future results.

This material has not been prepared in accordance with the legal requirements designed to promote the independence of investment research produced by broker-dealers or regulated financial institutions. This document is not a research report under FINRA, SEC, FCA, or MiFID II definitions. It has not been reviewed, endorsed, or approved by any regulatory authority, including FINRA, the SEC, or any similar body. Alpha Talon is not a registered investment adviser, broker-dealer, or financial institution under SFC or MPFA.

Readers should conduct independent research and due diligence before making any investment decision. You should consult a licensed investment adviser, registered financial professional, tax specialist, or attorney regarding your specific financial situation and risk tolerance. Nothing in this publication establishes any fiduciary relationship, advisory relationship, or obligation on the part of the author or Alpha Talon toward any reader.

The author and affiliated accounts may hold long or short positions in the securities and financial instruments discussed in this report and may trade in them before, during, or after publication without further notice. These positions may be contrary to the views expressed herein. The author does not receive compensation from the issuers of any securities mentioned. Alpha Talon does not have investment banking relationships, commercial relationships, consulting arrangements, or compensation agreements with the companies discussed in this document.

No part of the author’s compensation is directly or indirectly related to the specific recommendations, analyses, or opinions expressed in this report.

All investments involve risk, including loss of principal. Certain securities discussed may be speculative or volatile and may not be suitable for all investors. Clinical trial failures, regulatory decisions, market conditions, macroeconomic shifts, geopolitical developments, and competitive pressures can significantly impact the securities analyzed. This information is provided “as is,” without warranty of any kind, express or implied.

Securities mentioned herein are not guaranteed, not insured, and not protected by SIPC except as applicable for brokerage custody, and are not obligations of, or guaranteed by, any bank or government agency.

Alpha Talon, its author(s), and affiliates expressly disclaim all liability for errors, omissions, or any direct, indirect, incidental, or consequential losses arising from the use of this material. Use of the information is at the reader’s sole risk.

This material may not be distributed or used in any jurisdiction where such use or distribution would be contrary to local law or regulation. Readers are responsible for complying with applicable securities laws.

Disclaimer, Disclosure, Conflicts & Copyright Notice

Investing in securities involves significant risk, including the possible loss of principal. Equity investments may fluctuate in price, sometimes dramatically. Securities —such as those discussed herein—are subject to heightened levels of consumer, macroeconomics, competitive, and operational risk. Forward-looking statements, projections, price targets, valuation scenarios, and estimates included in this report are inherently speculative, based on numerous assumptions, and may differ materially from actual outcomes. Past performance is not indicative of future results.

No part of the author’s compensation is directly or indirectly related to the specific recommendations, analyses, or opinions expressed in this report.

Alpha Talon Investment Research

Discussion about this post

Ready for more?