Reflexivity and the AI Boom: A Sorosian Analysis

After sitting on my desk for months after borrowing a copy from my colleague Kenneth Beanland – my 11-week-old was begging me for some more interesting nighttime reading so we have been working our way through George Soros’ The Alchemy of Finance. Image at the end.

The book provides an intricate exploration of how financial markets are shaped not just by traditional economic theory, but by the unpredictable interplay of human psychology, feedback loops, and shifting global forces. Soros challenges the dominant belief that markets naturally seek price equilibrium or reflect “true” value. Instead, he introduces his iconic theory of reflexivity – where market participants’ perceptions and biases actively alter the fundamentals, which then further impact those perceptions. Soros describes the market as a dynamic system, constantly oscillating through networks of positive and negative feedback, where prices and narratives co-create reality. He illustrates through real-time trading diaries, global case studies, and philosophical musings that the market is more laboratory than machine – a place where hypotheses are tested and sometimes discarded as quickly as the rules seem to shift.

Throughout the book he reviews reflexivity in real time as well as its involvement in historic market environments – some of these periods include the 1960s, 1990s and early 2000s.

In the 1960s, conglomerates discovered a powerful mechanism: they could use their elevated valuations as currency to make acquisitions, which would boost earnings per share – they could buy growth and be rewarded rather than having to organically generate it. Investors, focused myopically on total earnings growth rather than its source, rewarded these acquisitions with even higher price-to-earnings (PE) multiples. This created a self-reinforcing cycle. High stock prices enabled more acquisitions, which drove further multiple expansion, which provided more acquisition currency. The prevailing bias was clear – conglomerates could receive high multiples “just for promising to put it to good use in making acquisitions”. The reflexive loop continued until reality could no longer sustain expectations – until conglomerates ran out of sufficiently large acquisition targets and the accounting manipulations became transparent.

In the late 1990s dotcom bubble, Soros’ theory of reflexivity played out through a similar feedback loop, but with technology and internet stocks at the centre. Investors, venture capitalists, and entrepreneurs collectively believed that the digital future would deliver unlimited growth and profitability – regardless of fundamentals. The prevailing bias here being – “Internet companies are revolutionizing everything—growth and valuations can only go up.” As money flooded into tech stocks, sky-high valuations enabled companies to raise more capital, spend lavishly, and launch ever-more ambitious ventures. These actions seemed to validate the market’s optimism, attracting even more investment and bidding prices higher. Positive feedback between rising share prices, easy capital formation, and widespread media enthusiasm drove the market to unsustainable extremes.

But this optimism distorted reality, separating expectations from actual earnings or sustainable business models. When it became clear that many companies couldn’t deliver profits, the narrative shifted. Perceptions reversed: capital dried up, valuations collapsed, and the market underwent a brutal correction.

In the early 2000s, banks, investors, and homeowners fed the US housing bubble through a powerful reflexive mechanism. Cheap credit and rising home prices reinforced each other: lenders extended more mortgages as prices climbed, validating optimism. Wall Street packaged risky loans into securities, channelling global capital into real estate. The prevailing bias was clear – housing prices “could only go up.” Homeowners borrowed more, feeling wealthier; investors speculated, fuelled by easy money and belief in endless gains. Each party’s perception of safety directly shaped rising market conditions. When defaults rose and prices fell, the loop reversed – credit contracted, asset prices crashed, and losses rippled through the system.

Just as Soros observed in the conglomerate boom of the 1960s, the 90s Dotcom Bubble & later in the 2000s US Housing Bubble – we are witnessing an eerily similar reflexive pattern unfolding in today’s artificial intelligence infrastructure buildout. The parallels are striking, though the scale is unprecedented.

The AI Infrastructure Boom: A Modern Reflexive Cycle

Today’s AI boom exhibits the same reflexive structure, but with circular financing replacing acquisition-driven growth. The hyperscalers; Microsoft, Amazon, Alphabet, and Meta – are deploying unprecedented amounts of capital – an estimated $300-400 billion in 2025 alone. This represents ~70% of their operating cash flows being redirected into AI infrastructure.

The self-reinforcing process operates through interconnected deals that blur the line between genuine demand and manufactured growth. Bloomberg recently posted this graphic to help visualise the incestuous nature of the deal flow and overarching demand.

The Prevailing Bias and Market Perception

Just as conglomerate investors in the 1960s focused on earnings growth regardless of quality, today’s market participants are fixated on AI infrastructure spending as a proxy for future dominance. The prevailing bias is that companies must spend massively or risk falling behind in the “AI arms race”. This creates what Soros would acknowledge to be a reflexive loop in action – market participants’ perceived urgency to expense unprecedented amounts of capital on this infrastructure build out provokes other participants to follow suit – in turn influencing the market’s reality.

The stock market rewards this spending lavishly. The hyperscalers have seen parabolic returns in recent years with AI companies accounting for 80% of US stock market gains in 2025. The cohort is expected to spend 30-50% of revenue on capex this year – levels last seen from AT&T and CISCO spending 25% and 14% at their respective peaks in the late 90s. early 00s. These elevated valuations provide cheaper financing costs (via equity deals & low coupon bonds), attract exceptional talent, and create a wealth effect that validates further spending.

Source: Bloomberg

The Manipulation Function

In Soros’ theory, the manipulative function describes how participants’ actions influence the underlying fundamentals. The AI boom demonstrates this powerfully:

Capital availability creates its own validation. Hyperscalers are spending 100s of billions in a single quarter on AI infrastructure. This massive deployment creates artificial/perceived scarcity for GPU capacity, driving prices higher and creating the perception of insatiable demand (despite GPU lending costs declining markedly). Nvidia reported that its top two customers accounted for 39% of Q2 revenue – a dangerous concentration that reveals how much of this “demand” comes from a handful of circular financing arrangements.

The solution? Give some other players money to diversify their customer base..

The revenue reality diverges dramatically from the spending. While $300-400 billion flows into AI infrastructure annually, actual AI revenue tells a different story. OpenAI projects $13 billion in revenue for 2025 – impressive growth from $4 billion, but minuscule compared to the infrastructure being built. Anthropic projects $9 billion annualized revenue by end of 2025. The entire generative AI software and services market is only expected to be $20bn – $30bn annually in FY25 and < $30bn in the next 12 months.

Goldman Sachs estimates that hyperscalers need AI productivity gains worth $8 trillion in present value to justify current investment levels. Morgan Stanley projects GenAI revenue could reach $1.1 trillion by 2028, up from $45 billion in 2024. Yet even these optimistic projections reveal a fundamental imbalance: analysts estimate the AI infrastructure buildout requires roughly $150 billion in annual revenue just to break even on depreciation and operating costs, while current AI-specific revenue is a fraction of that. There is also a growing concern over the depreciation schedules guided by the majors with Nvidia’s 1 year product cycle putting Moore’s Law on notice and bringing into question the relative useful life of high-octane compute.

The Self-Reinforcing Loop

The boom-bust sequence operates through distinct phases:

Phase 1: The Initial Trend (2022-2024)

ChatGPT’s November 2022 launch created a genuine underlying trend—generative AI demonstrated real capabilities. This sparked legitimate investment as companies recognized transformative potential. The potential for this technological innovation is absolute (it even helped write this article), but the math ain’t mathing…

Phase 2: The Misconception Develops (2024-2025)

The misconception that emerged: unlimited spending on infrastructure will automatically translate to proportional revenue and earnings growth. Market participants began conflating investment with demand. To date this has not been the case. In fact, it’s been largely capital destructive. If you exclude the chip manufacturers (the major winners of this thematic – NVDA/AVGO) the Hyperscalers are spending north of 160% of all revenue growth and 175% of EBITDA growth on CAPEX. Put another way, if you assume that ALL the revenue & EBITDA growth the Hyperscalers are generating can be ascribed to the CAPEX (certainly not the case) the best outcome across the cohort is AMZN who this year will get 60c back in revenue growth for every $1 spent (most are below 50c) – a loss of 40c/$1 spent. Looking at EBITDA – the cohort gets between 20 – 57c back in EBITDA growth. While you can make the argument that this will be recouped in future years, the issue I see largely stems from the rapid depreciation of the CAPEX (reducing the payback period timeline) and the growing OPEX likely to be incurred to run these increasingly demanding datacentres.

Source: Bloomberg

Phase 3: The Positive Feedback Loop (Current)

Each new spending announcement reinforces market confidence, driving stock prices higher, which lowers capital costs, enabling more spending. JPMorgan notes that AI-related capex contributed 1.1% to GDP growth in the first half of 2025, exceeding consumer spending’s contribution. This creates a macroeconomic feedback loop where AI spending props up economic growth, which validates more AI spending – and likely brings the US Administration into the fold as they look to perpetuate the spending and maintain the perception of strong headline growth.

The circular deals amplify this: when Nvidia invests in OpenAI, which buys Nvidia chips, Nvidia can report “rising sales,” OpenAI can claim it’s “growing,” and investors conclude the sector is booming.

Phase 4: Market Structure Changes (Emerging)

Just as the conglomerate boom created a specialized investor class, AI has spawned an entire ecosystem: AI-focused venture capital, “sovereign AI” government investments, specialized “neocloud” providers, and investment vehicles dedicated solely to AI infrastructure. The market structure itself has been reorganized around the theme.

The Approaching Reversal

In reflexive systems, the boom continues until fundamentals can no longer support perceptions. Several factors suggest we’re approaching this inflection point:

The Revenue Gap Is Widening. Despite 50%+ annual growth in hyperscaler spending, AI revenue growth, while still impressive in percentage terms, is markedly slowing (20-30% to 10-15% YoY). OpenAI is burning $8 billion annually while generating $13 billion in revenue. Even at projected 2026 revenues of $30 billion, OpenAI would need years to justify the infrastructure being built around it.

Production & Operational Constraints Are Real. There are rising concerns over the access to energy and water for cooling required to operate the growing number of data centres required.

The Competitive Dynamics Are Self-Defeating. When tech giants offer AI features as loss leaders subsidized by profitable core businesses (I am sure you have seen the recent price cuts/free trial extension to ChatGPT), standalone AI companies cannot charge premium prices. This creates a race to the bottom in monetization even as infrastructure costs soar. ChatGPT has 700 million weekly active users but generates only $12-13 billion in annual revenue – barely $17 per active user per year.

The Bust Sequence

When the reversal comes, the reflexive loop will operate in the opposite direction with equal force:

Declining stock prices will increase capital costs, making further infrastructure investment harder to justify. As multiples compress, the “currency” for these circular deals deteriorates. Problems swept under the rug during the boom – divisional operating losses, concentrated customer bases, dubious revenue recognition – will consume attention and bias investors against AI infrastructure plays.

The feedback loop reverses: lower valuations → reduced spending → overcapacity → price competition → lower revenues → further valuation compression. The very market structure that amplified the boom, will amplify the bust.

Why This Time Follows The Same Pattern

Some argue this cycle is different – that AI genuinely transforms productivity and therefore justifies any level of investment. Yet this was precisely the argument during the conglomerate boom (diversification creates value), the dot-com bubble (the internet changes everything), and the 2008 housing bubble (real estate always goes up).

The technology may be revolutionary, but the financing structure is reflexive. When Microsoft can afford to spend $80 billion on AI infrastructure as it’s considered essential to “preserve their monopolies”, when Nvidia’s growth depends on investing in its own customers, when projected revenues are orders of magnitude below required break-even levels – these are the hallmarks of a reflexive, boom-bust process in motion.

The principle of fallibility operates throughout: market participants believe unlimited AI spending will create proportional value, and this belief drives actions that temporarily validate it through circular financing and market enthusiasm. But the underlying fundamentals – actual revenue from actual customers paying for actual value delivered, remain stubbornly disconnected from the spending levels.

As Soros initially published in 1987 – financial markets are self-destabilizing, and boom-bust cycles are endemic to the financial system. The AI infrastructure boom exhibits all the classic signs; a real underlying trend, a prevailing misconception about its implications, circular financing that manufactures apparent validation, market structure reorganization around the theme, and a growing divergence between perception and reality.

The question is not whether this reflexive cycle will reverse, but when – and how severe the resulting bust will be when reality finally reasserts itself. Unlike a game of pass the parcel, you don’t want to be the one stuck holding the bag when the music inevitably stops – until then enjoy the party.

LD (& My Co-Author)