East-West or Bust: CoreWeave, Oracle, Google?
How Many AI Investors Really Understand What They Are Investing In?
Every major technological revolution has a moment when people mistake the visible surface for the deeper shift. During the railway boom, it was trains rather than steel; during the dot-com era, websites rather than fiber. Today, it’s AI models rather than the radical re-architecture of compute happening out of sight.
The difference now is scale: the infrastructure being built for AI is so vast, so capital-intensive and so structurally different from the past that it will redraw the maps of power across the entire technology sector.
AI Is Moving East-West
Here’s the issue that many haven’t grasped. The cloud infrastructure (data centers) of yesteryear, built by today’s hyper-scalers, were not optimized for AI.
Traditional cloud is mostly outside-in traffic: lots of small requests coming from users’ phones/laptops into the data center, with answers going back out. In networking terms that’s known as “north‑south”: data moving mainly between the data center and the outside world.
Yet with modern AI training, the heavy traffic is “side‑to‑side” inside the building. Huge numbers of GPUs (specialized chips) are talking to each other constantly to stay in sync while they train a big model, so most of the data moves between machines inside the data center, not in and out of it. That is termed “east‑west” traffic, referring to movement between GPU clusters.
This explains why so much CAPEX spending is taking place around AI. The data centers of tomorrow need an architecture that is entirely different to the cloud infrastructure of the past. Said differently, AI doesn’t just “stress” existing data centers, it forces them to be redesigned around these GPU swarms and their internal traffic patterns: from how networks are wired and managed, to how capacity is planned and where money gets invested.
Data centers are shifting from being optimized for lots of small web requests to being optimized for giant, tightly synchronized AI jobs.
This internal traffic needs to work well:
Ultra‑low latency: the chips must “hear back” from each other almost instantly, or the whole training run slows down while fast GPUs wait for slow ones to catch up.
Massive bandwidth: the connections between machines need to be extremely “wide pipes” (hundreds of gigabits to terabits per second) so all that data can move without creating a queue.
Lossless transmission: if data packets are dropped and resent, it stalls all the GPUs waiting for that piece, wasting very expensive hardware time. For these jobs, even small loss is unacceptable.
Dynamic adaptability: the pattern of who talks to whom can change suddenly, so the network has to be smart and automatically reroute and rebalance traffic to avoid hotspots instead of being fixed and static.
Whereas the hyper-scalers captured an oligopoly in terms of ‘North-South’ data centers, the need for ‘East-West’ cloud infrastructure has opened the door for others to compete.
Crucially, this is not merely an expansion of the existing cloud infrastructure; it is the construction of a new, parallel compute architecture. The traditional cloud was built on a CPU-based, serial compute model, optimized for cost-down efficiency and generalized utility. The new AI infrastructure, by contrast, is a dense form of parallel compute (epitomized by specialized GPUs and custom accelerators), optimized for pure performance at any cost.
Companies like [$CRWV] CoreWeave, [$NBIS] Nebius and even [$ORCL] Oracle (which was previously considered a laggard in the cloud era) are now emerging as key players in servicing this new architecture. They differentiate themselves by being GPU- and networking‑first (often bare metal, tightly coupled clusters, aggressive preemption, and flexible placement) rather than general-purpose virtualized clouds; this better matches large-scale training and high‑throughput inference.
AI is reshaping compute architectures, pushing the industry toward disaggregated, accelerator-centric designs that require enormous east-west bandwidth, lossless fabrics and far smarter scheduling. That shift is forcing today’s hyperscalers to rethink their networks, pricing models and platform abstractions, or risk losing the most profitable AI workloads.
But does that mean investors should sell out of the hyperscalers and rotate into the AI-native cloud infrastructure players? Absolutely not. The hyperscalers remain investment-grade cash machines and, crucially, they can largely fund the required CAPEX internally. By contrast, CoreWeave, Nebius and even established players such as Oracle are spending aggressively and burning through cash at a pace that outstrips what they can generate. That strain is now showing up in their credit risk profiles, most visibly in the widening of their CDS spreads (the market’s way of pricing the cost of insuring their debt against default).
Take Oracle. Despite its investment-grade rating, its CDS spreads have climbed to levels not seen since the 2008 financial crisis. Credit agencies are increasingly uneasy about projections that its net debt could triple in the coming years, putting meaningful stress on the balance sheet. Moody’s has flagged “significant” risks, and JPMorgan has already downgraded the company’s credit. Investors are watching closely to see whether Oracle can generate the returns needed to service this rising debt burden.
Oracle’s recent profit and revenue outlook missed estimates and it announced that its spending would increase by another $15 billion compared to prior estimates. Oracle shares fell more than 11% in after hours trade on 10th December as a result. It’s been a wild ride this year for Oracle investors, but over the course of 2025, it hasn’t outperformed the broader S&P500:
CoreWeave presents a different, but equally revealing, case study. The company went public in April and quickly captured investor attention thanks to its privileged access to NVIDIA’s cutting-edge GPUs, access reinforced by NVIDIA’s own equity stake. With Microsoft accounting for 62% of annual revenue, CoreWeave is operating with considerable customer concentration risk, yet that didn’t stop its stock from soaring more than 200% within three months of the IPO.
The problem is that many of these high-growth AI infrastructure firms are still trying to run the old “growth at any cost” playbook. That strategy thrived when money was cheap and investors were willing to subsidize endless expansion. It’s far less compelling in a world defined by higher rates, persistent inflation, and a renewed focus on cash flow. What works for an asset-light SaaS company simply doesn’t translate to hardware-heavy, capex-intensive operators.
CoreWeave’s own CDS spreads now sit in territory typically associated with distressed credit. Markets are clearly uneasy about the company’s deeply negative free cash flow and its reliance on continuous debt issuance and equity raises to fund growth. There’s also the uncomfortable reality that its core collateral (NVIDIA GPUs) can become technologically obsolete at a breakneck pace.
December 2025, CoreWeave announced plans to raise another $2 billion via convertible notes, so the capital burn continues at pace.
Nebius is not much better positioned. It remains deep in investment mode, burning cash and leaning on new debt secured against future contract revenue. For now, AI optimism is keeping the funding taps open. But if market sentiment turns, companies operating with this level of financial strain could find themselves in serious trouble.
It’s not hard to imagine a scenario where hyperscalers ultimately benefit from the current spending boom by picking up these businesses at distressed valuations. If that happens, investors paying steep multiples today for exposure to fashionable AI names may find themselves on the wrong side of the trade.
CoreWeave’s recent missteps haven’t helped sentiment. Its failed attempt to acquire Core Scientific, an all-stock $9 billion deal, was widely seen as an attempt by CoreWeave to use its over valued equity as currency to acquire growth. The problem was that the two companies shared little beyond similar names. Core Scientific, fresh out of bankruptcy (January 2024), is one of North America’s largest Bitcoin miners, operating power-dense data centers built for crypto workloads. CoreWeave, by contrast, is scrambling to meet skyrocketing demand for GPU capacity from enterprise and generative-AI clients.
The merger logic was straightforward: repurpose Core Scientific’s footprint to accelerate CoreWeave’s AI cloud expansion, while giving Core Scientific an exit from the volatility of Bitcoin mining. CoreWeave was already Core Scientific’s largest customer, with $8.7 billion in hosting commitments over 12 years, and acquiring it would have eliminated $10 billion in lease payments, secured 1.4 GW of power capacity, and delivered an estimated $500 million in annual cost savings by 2027.
But the market didn’t buy it. Backlash from both sets of shareholders forced the companies to abandon the deal. CoreWeave’s stock promptly slid from $131.06 at termination to $79.36, a clear signal that without Core Scientific’s infrastructure to rely on, investors are re-evaluating the true cost and complexity of scaling its GPU cloud.
In short, predicting the winners in this race is far from easy.
From Specify to Verify: Understanding Artificial Intelligence
What exactly is Artificial Intelligence?
As Albert Einstein and Richard Feynman used to say, “If you can’t explain it, you don’t understand it.”
Most people don’t understand it and couldn’t explain it, but that doesn’t stop them investing.
In traditional computing, a human writes clear rules that say, “When you see X, do Y,” so the computer just follows instructions. If the input is the same, the output will always be the same. The rules are fixed, and the computer simply follows them.
Artificial Intelligence, on the other hand, takes a different approach. Instead of specifying exact instructions, we let the computer “figure it out” by trying different actions and scoring the results. Think of it like playing a game of “hot and cold” as a child when searching for something your parents hid. As you get closer to the hidden object, you’re told you’re getting “hotter,” and as you move farther away, you’re told you’re “colder.” Over time, you adjust your approach based on these cues and eventually find the treasure.
In AI, this process of trial and error, where the computer continuously tests different solutions and adjusts based on feedback, is called machine learning. The key difference from traditional computing is that rather than writing fixed rules, we focus on verifying outcomes and allowing the system to learn from its experiences. It’s an exploratory and probabilistic approach, one where the system continuously fine-tunes itself to optimize its performance.
In essence, AI shifts from a world of rigid specifications to one of dynamic verification and continuous improvement. This is what is meant when people speak of ‘training AI models’.
If you haven’t already seen it, I highly recommend watching this documentary about Nobel Prize winning Demis Hassabis, who developed Deep Mind, now part of Alphabet (click to view a quick trailer):
Is AI Exuberance Justified?
The current era of Artificial Intelligence is fundamentally rooted in the emergence of transformer models in 2017, a development out of [$GOOG] Google (trading as Alphabet) that provided a new, highly scalable method for building AI.
Since that inflection point, and particularly following the widespread adoption catalyzed by the launch of ChatGPT in late 2022, the industry has been dominated by the relentless upward trajectory of scaling laws, which dictate a continuous, predictable improvement in model capability correlated directly with increases in training data, computational power, and model size (parameter count). The sustained validity of these laws is the single most important driver of the modern AI landscape.
The core insight underpinning the current AI surge is the non-linear relationship between resources and output. The initial scaling laws observed were statistical relationships, showing that as compute, data, and parameters grew, the model’s performance, technically measured by a falling test loss ratio, consistently improved. This principle has held across multiple generations of models, validating the enormous investments being made.
However, more recently, the industry has seen the arrival of a second scaling law: test-time scaling. This observation demonstrates that performance improves not just with the size of the model, but with the duration of its “thinking,” or the number of inference cycles it goes through. Bigger models are better, but crucially, models that think longer are also better. This dual dynamic has profound implications, suggesting that the path to Artificial General Intelligence (AGI) and advanced agentic AI is tied directly to the willingness and ability to deploy ever-greater computational resources.
The speed of this progress is what truly distinguishes the AI revolution from prior technological cycles. Historically, Moore’s Law, the driver of the information technology revolution since the 1950s, delivered an order of magnitude improvement roughly every decade. In contrast, contemporary AI models are improving at a rate of somewhere between half and a full order of magnitude every single year.
This radically steeper slope creates an environment that is both exhilarating for investors in the infrastructure layer and deeply disconcerting for legacy businesses. The statement that “the models you use today are the worst they will ever be” captures the essence of this extraordinary and persistent rate of change, making it exceedingly difficult to imagine the downstream effects and applications of future, vastly more capable systems.
GPT: General Purpose Technology and Tokens
The most useful framework for understanding AI’s economic impact is to view it as the next General Purpose Technology (GPT), following in the footsteps of the internet, electricity, steam power and steel. A GPT must possess three key characteristics: it must be pervasive, it must continue to improve and it must facilitate spillover innovations across the entire economy.
So far, AI appears to satisfy all these conditions. Its pervasiveness is already evident in the rapid adoption: ChatGPT reached 700 million weekly active users in a market already saturated with tech giant, demonstrating an extraordinary compression in the adoption timeline compared to previous GPTs (which took decades or generations).
However, this rapid technological advancement brings a profound and disconcerting economic feature: deflation. The speed at which the cost of providing AI services has fallen is phenomenal. The cost of one million tokens1 using GPT-4 initially cost approximately $50; for the superior GPT-5 model, that cost has plummeted to around 15 cents.
Google is leading the charge on driving token prices lower. It can price AI tokens cheaper mainly because it owns the full stack (chips, data centers, models, distribution). Its Tensor Processing Units (TPUs), like v5e, are explicitly designed for cost‑efficient inference (lower power, lower system cost, dense networking).Then, because Google designs much of its server, networking and datacenter stack in‑house, it can reduce overheads like networking and system integration, which are now a meaningful share of training and inference costs for large models. In addition to all of these advantages, it can cross‑subsidize AI from higher‑margin businesses (search ads, Workspace, Android, YouTube) and so run on far tighter margins, or even negative margins, simply to grow market share and apply pressure on competitors.
The market is certainly starting to back Google in the AI race. The following chart compares the performance of two baskets of stock: one contains the companies most closely aligned to OpenAI infrastructure, while the other comprises companies more closely aligned to Alphabet AI. Both were running side by side until Google released Gemini 3; then the market expressed its view very clearly.
Token pricing is part of a broader share‑grab: Google is positioning “Flash”‑class models as the cheapest adequate default for many applications, forcing others either to match pricing or to justify a quality premium.
This is good news for the consumer of AI as consumption has grown exponentially, but such a rapid level of deflation can destabilize business models that rely on predictable pricing. The short-term challenge is navigating the part of the technology cycle where costs fall before the commensurate revenue response materializes.
The long-term answer to this deflationary pressure lies in Jevon’s Paradox: as the resource (compute/AI capability) becomes cheaper, demand for its use will explode. This pattern has been observed in countless cycles, including bandwidth during the internet boom and the cost of electricity. The challenge is not whether demand will materialize, but surviving the period where the cost of the resource is collapsing rapidly, forcing participants to make high-stakes, forward-looking investments based on a future expectation of demand explosion, not current revenue figures.
The Mammoth CAPEX Cycle
The conviction that scaling laws will hold and that AI is a true GPT has triggered an unprecedented capital expenditure (CAPEX) cycle, driven primarily by the world’s largest companies: the hyper-scalers. Annual CAPEX estimates are now soaring, triple the levels seen just before the launch of ChatGPT, pointing to a massive, systemic reallocation of capital towards compute.
The motivations for this gargantuan spending fall into three categories, which collectively justify the risk of hundreds of billions in investment:
Defensive Spending (Existential): For dominant players like Google, AI is inherently disruptive. The rapid loss of market share in online query activity is an existential threat, presenting a terminal value problem for its core search business. The price an incumbent will pay to protect quasi-monopoly rents is essentially infinite.
Offensive Spending (ROI): The immediate returns are significant and growing quickly. Companies like [$MSFT] Microsoft have seen billions in AI-related revenue, and [$META] Meta projects tens of billions in uplift from better targeting and optimization driven by AI. While these numbers alone may not yet justify the total CAPEX, they demonstrate rapid monetization potential in the near term.
Optionality Spending (The Race to AGI): This is the hardest factor to quantify but arguably the most potent driver. Leaders view the pursuit of AGI, the point where a machine can perform at human or better-than-expert levels across all cognitive tasks, as a race where the winner takes a prize valued in the trillions or tens of trillions. As Mark Zuckerberg, CEO of Meta noted, missing out on AGI is a far greater shame than misspending one or two hundred billion dollars. This desire to secure a leading position in super-intelligence and agentic AI explains the urgency and the willingness to take on massive capital risk.
The Accounting Issues
The way companies account for their AI spending matters far more than most investors realise. In a heavy-capex technology cycle, the mix of accounting choices, cost assumptions, incentives and plain old corporate psychology can reshape the economic picture.
Right now, firms pouring billions into advanced GPUs are stretching depreciation schedules in ways that deserve scrutiny. Hardware that was typically written down over three or four years is suddenly assumed to last six. Given the extraordinary cost of these chips, that shift has a meaningful impact: longer depreciation lowers quarterly expense, lifts reported earnings and flatters the apparent returns on AI investments. Shorten the schedule and those returns would shrink fast.
The question is whether these assets genuinely deliver six years of productivity, or whether companies are simply massaging their numbers. If the hardware ages more quickly than the accounts suggest, real margins are thinner than reported and eventually the shortfall will show up in write-downs, profit warnings and sharp market corrections.
With hardware performance doubling at a relentless pace (Moore’s Law), it feels counterintuitive, if not outright optimistic, to double the depreciation period . At the very least, it raises the uncomfortable possibility that companies are obscuring economic reality.
Incentives complicate matters further. Pay management to grow revenue at any cost and that’s exactly what they’ll do, even if it burns through capital. Tie compensation to adjusted metrics like earnings and you create a fresh incentive to engineer away inconvenient truths.
In reality, many of these AI infrastructure companies are investing simply to stay in the game, rather than decisions being based on sound economics. Return on Invested Capital? Kick that concern down the road!
NVIDIA’s recent commentary feeds into this debate. On its earnings call, CEO Jensen Huang argued that A100 GPUs shipped six years ago are still running at full utilization thanks to steady improvements in the CUDA software stack. The message is clear: NVIDIA’s accelerators may have a longer economic life than critics assume. Naturally, the company benefits from making that case, especially at today’s price points. Still, if even partially true, it would help to explain why depreciation schedules are being stretched.
Optimism may not be the only force at play, necessity plays a role too. When hardware is cheap, replacing it is easy to justify; when it’s extraordinarily expensive, as cutting-edge GPUs are, the incentive is to squeeze every last year out of them. There is precedent for this kind of ‘freakonomics’. In the U.S. the average age of a car in the 1970s was 5.6 years, yet today it stands at 12.7 years. The rising average cost of new and used cars explains this phenomenon as the two factors are very strongly correlated. In the 1970s a new car represented about 40% of the median annual family income, today it stands at 60%.
In relation to GPUs, both explanations are plausible, yet neither is easy to test. We simply won’t know until it’s too late.
Layered on top is another potential risk: vendor financing. If hardware manufacturers are effectively funding customers who can’t otherwise afford their products, the system becomes a fragile chain held together by its weakest participant.
There’s precedent here, and the parallels are worth remembering. IBM, once the world’s most valuable company, drifted into financial engineering during Lou Gerstner’s tenure in the 1990s. One of his tricks was vendor financing. Gerstner relied heavily on customer financing through vehicles like IBM Global Financing, which securitized loans and moved them off IBM’s balance sheet. IBM booked the full revenue and profit upfront, even though it was effectively subsidizing customers through cheap credit. The top line looked great, returns on capital looked better, but the economics were far less robust than they seemed. The market eventually caught up with the shenanigans and the IBM valuation tumbled. Today IBM is a shadow of the company it once was.
Against that backdrop, consider the allegations that NVIDIA has invested in special-purpose vehicles linked to major AI customers including CoreWeave and OpenAI, where those funds are used to buy NVIDIA’s own GPUs2.
History doesn’t repeat, but it does have a habit of rhyming!
Market Structure and the Private/Public Divide
While the infrastructure build is largely a hardware cycle playing out in the public markets, with NVIDIA being the undisputed standout success story so far, much of the core intellectual property development is occurring within private labs like OpenAI, Anthropic and XAI. This structure diffuses the signals available to public investors. However, the sheer scale of the capital needed for the next generation of models (Grok 4, GPT-6) is so large that it is moving beyond the capacity of even the deepest venture and sovereign wealth funds.
The non-linear, capital-intensive nature of the scaling laws suggests that public markets will eventually be required to meet the estimated $3 trillion of AI investment needed through 2028. Importantly, the source of this capital is fundamentally different from the dot-com bubble of the 1990s: half comes directly from the hyper-scalers themselves, which are highly cash-generative investment grade businesses. This inherent financial stability suggests that while many draw parallels to the “dot-com bubble”, the foundations are far more robust than those built on the highly leveraged balance sheets of new, unproven telcos two decades ago.
The Investor’s Valuation Conundrum
The arrival of AI poses a direct and immediate terminal value risk to companies whose moats were built on proprietary data or legacy infrastructure. Investors are increasingly applying an “AI lens” to portfolios, looking at how companies will fare in an “AI-first, agentic world.”
The effects are already visible: fantastic, storied businesses like [$IT] Gartner or [$WKL] Wolters Kluwer, with high ROIC and great histories, are seeing their moats, once defined by proprietary data, being chipped away by highly capable, generalized models. Capital that would have flowed into these legacy industries is now being reallocated to AI.
For the hyper-scalers, the challenge is navigating the transition from the old to the new architecture. [$AMZN] Amazon Web Services (AWS) was the brilliant conduit for the cloud, but the assumption that it would automatically dominate the parallel compute era is no longer clear.
Existing cloud leaders must balance continued maintenance CAPEX in the back book of the old serial compute, necessary to protect their lucrative existing business, with growth CAPEX on parallel compute to compete in a brave new world.
Alphabet’s situation is the most acute. While possessing every advantage: chips (TPUs), data, distribution and talent, its core problem is cannibalization. It must transition from the dominant, highly profitable economics of its traditional search business to a more contested agentic model. Although the company is being forced to ship product and take risks, a positive outcome of the competitive pressure, the threat remains that the search business may stagnate, becoming a legacy market that simply stops growing as all incremental consumer and commercial activity shifts to AI agents. The outcome of Google’s fight against OpenAI will ultimately depend on whether its cost-per-token advantage proves decisive, or whether the historical trend of winners continuing to win in consumer tech holds true.
When we consider the rise and fall of names such as IBM, Intel and Netscape, it becomes clear that size and market dominance are not themselves guarantees of future success.
The timeline for AI adoption is following a predictable S-curve: slow through the innovation phase, followed by a take-off phase and a rapid, accelerating adoption phase. Cohort analysis shows that the longer people use ChatGPT, the more they use it, accelerating the rate of adoption.
But there is also a hype-cycle
We’re now deep into the peak of inflated expectations, where the companies at the center of the AI boom are being priced as if nothing can go wrong. Little attention is being paid to the sheer level of risk embedded in these valuations. It’s precisely in moments like this that capital is most easily destroyed, just ask the early railroad investors who learned that lesson the hard way.
A more modern parallel comes from the late 1990s. After the 1996 Telecommunications Act deregulated the U.S. telecom sector, companies launched into an all-out race to lay fiber across the country. Executives were convinced that internet traffic would double every hundred days, a claim that was later thoroughly debunked. They eye-watering amounts of money into massive long-haul fiber buildouts.
But the traffic never came close to those projections. Instead of doubling every hundred days, it doubled roughly once a year. And then came Dense Wavelength Division Multiplexing (DWDM), a breakthrough that allowed a single strand of fiber to carry more than a hundred times the data. By 2001, less than five percent of all U.S. long-haul fiber was actually lit. The rest sat unused as bandwidth prices collapsed by about ninety percent.
The fallout was brutal. WorldCom, Global Crossing and others plunged into bankruptcy, and The Economist estimated that investors ultimately lost more than a trillion dollars in market value. Ironically, a decade later that same fiber formed the backbone of the modern internet, but the original investors never recovered.
The key for AI investors today is in getting the timing right. It may require overlaying these curves, because timing is everything.
The Application Frontier: From Coding to Agentic AI
A key market question remains: will foundation models become commoditized, or will the “winner-take-all” dynamic prevail? The belief held by the leading labs is that they will not be commoditized at the frontier.
While simple tasks like document summarization are already commodities, complex, high-value tasks, such as curing cancer or running a sophisticated trading agent, will always be constrained by the best-performing model. The race is therefore fought at the margin of capability; so long as scaling laws hold, the incrementally better model, even if only slightly superior, will win and justify its massive development cost. In stark contrast, destruction of capital by those that don’t win will be significant.
The earliest and most successful applications have been in domains perfectly suited to the technology, such as coding and customer service:
Coding is an inherently well-structured task, allowing developers to see direct, measurable productivity gains from tools like GitHub Copilot.
Customer service, being a major cost center with high employee turnover, offers clear, quantifiable ROI through deflection rates and cost reduction.
Beyond these early enterprise applications, the consumer side is demonstrating extraordinary velocity. Non-work use cases are growing faster than work use cases, a testament to consumers finding immediate and increasing value in the technology.
However, predicting the ultimate killer applications remains the hardest part of the equation. Just as no one could have predicted [$UBER] Uber or [$ABNB] Airbnb at the invention of the mobile phone, the true, game-changing markets that will emerge downstream of AI’s improvement curve are currently unknowable.
The future vision is one of an agentic economy, where users are not “chatting away with a bot” but interacting with highly capable, trusted partners that act on their behalf, advising, buying and transacting. This shift of the model from a simple chat interface to something resembling an operating system, complete with preferences and memory, is where the next generation of durable consumer moats will be built, transforming the role of cognition and knowledge work from a constraining factor on economic production to a highly scalable, abundant resource.
Agentic AI will only fulfil its potential if autonomous agents can actually transact, and that means moving money. If that becomes reality, the biggest beneficiaries might not be the AI developers or the hyper-scalers, but the payment networks that build and run the financial plumbing underneath. Aviation offers a useful analogy: airlines carried the capital burden, yet it was the payment networks that captured the steadier, higher-margin economics of global travel. Innovation doesn’t always reward the party that bankrolls the infrastructure that makes it possible. The same was true of the rail-roads. Agentic AI may follow a similar pattern.
Conclusion
The AI revolution of the 2020s will go down in history as an infrastructure build powered by the predictability of scaling laws. It is laying the groundwork for a generation of models that will reshape the global economy at a structural level.
The enormous capital being deployed today carries both defensive and offensive motives, but it remains an open question whether all of it is commercially rational.
In the natural world, death is central to life because it makes room for renewal and makes evolution possible. Each generation expends energy that becomes the foundation for whatever follows. Economics follows a remarkably similar pattern. Joseph Schumpeter called it ‘Creative Destruction’ and the AI revolution is accelerating this cycle at a pace we’ve never seen before.
Companies, investors and individuals are pouring in enormous amounts of time, money and energy to build what they hope will define the future. Yet history suggests that many of today’s efforts may ultimately serve as stepping stones rather than endpoints. The capital expenditure, the breakthroughs and the infrastructure will become raw material for the technologies and businesses that eventually dominate.
There’s something both sobering and energizing in that reality. The work done today may not guarantee success for its current contributors, but it does propel the broader system forward for the greater good of mankind. Just as past generations of technology cleared the path for today’s innovators, the efforts of today’s builders might simply be paving the way for tomorrow’s winners, even if those winners emerge long after the original pioneers are gone.
Tokens in AI are small chunks of data (often pieces of text) that a model reads and predicts over, rather than whole sentences or documents at once. In language models, your text is first broken into these tokens, processed internally as numbers, and then the model generates new tokens one by one as its output. A token is typically a word, part of a word (like “un-” or “-ing”), punctuation, or sometimes a short phrase, depending on the tokenizer used. Because tokens are finer-grained than words, a sentence will usually contain more tokens than words. Models have a maximum number of tokens they can handle at once, called the context window, which caps how long your prompt plus the model’s reply can be. Pricing and performance are also usually quoted “per token,” so more tokens mean more computation time and higher cost.
I really enjoyed reading your posts. Informative and interesting as always.
Realy strong analysis on the token deflation point. The GPT-4 to GPT-5 cost drop (from $50 to 15 cents per milion tokens) is staggering but the Jevons Paradox framing makes sense, demand will eventually catch up once these models become cheap enugh. The challenge for investors is that bridge period where costs collapse before revenue materializes, and distinguishing between companies that can survive that squeeze versus those burning capital too fast.