AI investment is not a bubble; the real capital cycle has just begun.

Not long ago, David George, General Partner of a16z, and Gavin Baker, Managing Partner and CIO of Atreides Management, engaged in an in-depth discussion on the "Macro Logic of the AI ​​Era." They placed the current AI wave in a mirror image of the 2000 dot-com bubble, comparing the idle capacity of the "dark fiber" era with the current reality of "no dark GPUs." They believe that this investment is not speculation, but a rational expansion led by companies with abundant cash flow and the strongest balance sheets. Funds are flowing into computing power, models, and energy infrastructure at an unprecedented rate, driven by a triple force of survival pressure, technological leaps, and long-term returns. ▍Comparison and Reality of the AI ​​Bubble

Over the past thirty years, the technology industry has experienced several dramatic cycles of capital and belief. With each new wave of investment, similar questions arise: Is this just another bubble? However, comparing the current AI investment frenzy with the internet and telecommunications bubble of 2000 reveals differences almost entirely in asset structure and returns. The 2000 bubble was essentially an overexpansion of telecommunications infrastructure. Its hallmark was "dark fiber"—fiber optic cables laid but not yet operational. At the height of the frenzy, approximately 97% of fiber optic cables in the US were in a "dark" state, a quintessential symbol of overvaluation and underutilized capacity. Today, there are no "dark GPUs." All GPUs are running at full capacity; from research papers to production lines, signs indicate that computing power is being expended. Companies are no longer financing illusory potential but are exchanging every chip for immediate computational returns. Valuations are also fundamentally different. At the peak of the bubble in 2000, Cisco's price-to-earnings ratio (P/E ratio) reached 150-180, while today core computing power companies are trading at around 40. While the pricing structure remains expensive, it is supported by real cash flow. More importantly, the return on investment (ROI) has not been overdrawn. Major GPU investors saw their ROIC increase by an average of about 10 percentage points after initiating large-scale capital expenditures, meaning that computing power expansion is still operating within a positive return range. The US currently has approximately $1 trillion in data center assets and plans to invest another $3-4 trillion in construction over the next five years. This figure, converted to inflation-adjusted money supply, surpasses the US Interstate Highway system built over 40 years. Such infrastructure density intuitively suggests an impending bubble, but consider another set of data: in the past 17 months, global token processing volume has increased 150-fold. In other words, usage and supply have increased simultaneously. Typical characteristics of a bubble are slowing investment and inflated demand, while this round of growth appears more like a positive feedback loop between supply and demand. Another characteristic of the dot-com bubble was the barrier to entry and the difficulty of distribution. In 2000, internet companies needed to build two systems simultaneously: the "website" and the "user" side, creating a two-sided network effect. The dissemination of AI tools, however, is directly built on existing cloud computing and internet distribution systems. A single API interface or webpage entry can allow hundreds of millions of users to access it instantly. This "lights-on" infrastructure characteristic makes the application's cold start speed, coverage, and penetration rate far higher than during the dot-com boom. The source of funding and payment ability are also important differentiators. Currently, the main investors in AI are not startups, but global giants with trillions of dollars in cash flow. These investors collectively have approximately $300 billion in annual free cash flow and approximately $500 billion in cash on hand. Based on an estimated cost of $40-50 billion to power 1 GW of data centers, the entire system has a "liquidity buffer" of approximately $800 billion, and this is still growing at a rate of $300 billion annually. This structure means that even with short-term profit mismatches, funds can still support long-term trial and error and structural upgrades. This situation is not merely a matter of financial endurance, but rather a form of "survival investment." Leading companies are no longer focused on marginal profits, but rather on ensuring they don't fall behind in core technology areas. Internally, this is a consensus of "winning even at the cost of bankruptcy." The relationship between capital expenditure and technological breakthroughs has thus shifted to "strategic losses," meaning short-term financial fluctuations are tolerable, solely for the sake of long-term control. Correspondingly, there are concerns about "circular transactions." On the surface, some companies engage in closed-loop transactions of funds and products, but this is not the traditional bubble-type "self-buying and selling," but rather competitive mutual investment between chips, models, and platforms. While substitutability of funds certainly exists, the transaction volume is far lower than the fictitious revenue during the bubble phase. The core driver of this phenomenon is not financial illusion, but strategic defense. In the computing power supply chain, the boundaries of competition are blurred: suppliers can be both partners and potential rivals. Therefore, today's AI boom does not possess the characteristics of the false prosperity of the 2000 bubble. Its capital investment is concentrated in the real aspects of computing power, algorithms, and energy; funding sources are highly concentrated and self-circulating; and market usage and infrastructure expansion maintain a dynamic equilibrium. What is truly alarming is the decline in capital returns after the gradual reduction of efficiency dividends, rather than the bursting of a speculative bubble. ▍Giants and Computing Power: The Underlying Logic of Rational Expansion

Understanding this wave of AI investment hinges on recognizing the dominant players and return structures behind capital expenditures. Unlike the grassroots startups of the early internet era, those truly driving GPU procurement and data center construction today are global technology companies with historically high cash reserves. These companies collectively possess approximately $500 billion in cash reserves and can stably generate $300 billion in free cash flow annually, forming a self-sustaining, nearly closed investment system. Under this structure, even if the ROI of a single project deviates in the short term, it will not trigger systemic risk. A typical example is that the cost of powering a large-scale data center is approximately $4-5 billion per GW, a capital expenditure that remains within the acceptable range for companies with ample cash flow. In other words, this is a "capital expenditure wave led by profitable companies," rather than a speculative wave driven by financing. The relationship between these companies is both cooperation and competition. The capital cycle between GPU investment and model training is misinterpreted externally as a "self-buying and self-selling" revolving financing. However, from an industry perspective, this cycle is a competitive balancing mechanism. The transaction between chip suppliers and model labs is essentially binding the ecosystem with investment and locking in market share through supply. While some alternative funding options do exist in the market, their scale is relatively limited and primarily reflects strategic positioning. The current core of competition is no longer between traditional chip manufacturers like AMD, Broadcom, Marvell, or Intel, but rather concentrated on the battle between Nvidia and Google's TPU. The advantage of TPU lies in its deep integration with its own model and cloud computing system, forming a vertical system encompassing hardware, algorithms, and services. Nvidia, through its CUDA software stack and system-level integration, has evolved itself from a "chip company" into a "data center company." It no longer simply sells chips, but rather an entire replicable computing factory. This structural transformation has brought about unprecedented consolidation in the history of the computing power industry. The alliance between Broadcom and AMD is attempting to build a "second system" for large cloud providers, based on open Ethernet, to counter Nvidia's proprietary network. Just as Google took three generations to mature its TPU, the failure rate of ASICs remains high, and there may be several high-profile cancellations of projects in the next three years. Once Google begins selling TPUs, this chip landscape will be further reshaped. From a system perspective, Nvidia's technological advantage lies not only in its hardware stacking but also in its comprehensive oversight of the entire stack architecture. GPUs, NVLink, high-speed Ethernet, and InfiniBand constitute its "computing network," achieving low latency and high throughput for large-scale deployment through hardware and software co-optimization. This gives Nvidia unprecedented pricing power—it controls not just chips, but the entire computing ecosystem. This full-stack control also explains the rationale behind the so-called "circular investment" in the market. While some chip sales do come from funding for AI labs, the underlying logic is not financial fabrication but a natural product of supply and demand dynamics. Google, with its DeepMind and Gemini, naturally forms a closed-loop competitive landscape. To prevent competitors from controlling the upstream chip supply chain, Nvidia chooses to bind downstream users through investment, cooperation, and prepayment—a reasonable defensive strategy, not a bubble. The mentality behind computing power investment also reflects the giants' understanding of the survival of AI. Insiders have described this mentality as "winning even at bankruptcy," meaning that at the strategic level, short-term profits give way to long-term control. AI is seen as a fundamental track concerning the company's fate; any falling behind could mean rewriting history. This differs from traditional profit-seeking capital and is closer to the Cold War-era struggle for "technological leadership." The robust return structure also stems from the immediacy of computing power usage. Unlike the "dark fiber" era, every GPU today is involved in model training and inference. The explosion of training tasks means that computing power is almost never idle. From technical papers to internal industry data, it reflects that "GPUs are melting"—that is, computing power utilization is approaching its limit. Against this backdrop, capital expenditure has not lagged behind demand, but rather has been driven by it in real time. For long-term investors, this infrastructure expansion led by cash-rich, technologically advanced companies signifies predictability and resilience. Even if some companies experience a decline in marginal returns after investing in future Blackwell generations, the overall ROI remains positive. To date, this round of AI investment has been a "positive return cycle," with no signs of overcapacity or valuation decoupling. Therefore, while the outside world still understands AI through a "bubble" framework, the real change in the capital market lies in the upgrade of its logic. It is no longer a cycle of "trading dreams for valuation," but a competition of "cash for computing power." Only companies that can consistently generate free cash flow are qualified to participate in this game. Future risks may still exist, but they stem more from the physical limits of computing efficiency and energy consumption than from a financial collapse.

▍Business Differentiation of Model Ecosystem and Application Layer

While computing infrastructure is gradually solidifying, the focus of the AI ​​industry is shifting from "building factories" to "applications." However, the landscape of the application layer is far from clear. If ChatGPT is the Netscape of the AI ​​world, then the entire industry is still in the early stages of the internet era, before Google was born and Facebook was still in school. Mature ecosystems and winners are still forming.

This stage of uncertainty makes the infrastructure layer seem more "secure." Suppliers of computing power, networks, and models can lock in long-term returns through economies of scale, while the application layer is in a period of high turnover and high attrition. Investors generally acknowledge that humility is a necessary virtue in the face of rapidly evolving model cycles. Predicting which company will become the "AI version of Google" in the early stages is almost impossible. It's worth noting that the innovative nature of AI differs from that of the early days of the internet. It may not be a "disruptive innovation," but rather a more "sustainable innovation." Data, computing power, and distribution—these three elements that constitute a competitive barrier—are precisely in the hands of existing tech giants. These companies not only possess vast data assets but can also raise capital to purchase computing resources, while directly reaching users through operating systems, search engines, and social media channels. Therefore, provided execution is effective and the strategy is correct, the existing "Big Seven" could very well continue to dominate in the new cycle. This sustainable structure also brings a subtle risk: if execution fails, the outcome could be similar to IBM's obsolescence. For these companies, AI is a matter of survival, not business expansion. It forces every company with global distribution capabilities to re-examine its organizational structure, engineering pace, and internal culture. From this perspective, AI is not a simple technological shift but a "reprogramming" of management and decision-making systems. Another major differentiator in AI business models lies in the redefinition of gross profit margins. In the past, SaaS companies consistently maintained high gross margins of 80% to 90%, but the scale principle of AI applications means that computing costs will continue to suppress gross margins in the long term. The computational power consumed in training and inference makes AI products more like "heavy service industries," with their cost structure irreversibly shifting from pure software to hardware-intensive. This doesn't mean that AI applications aren't a good business. The key lies in reinterpreting how profits are distributed: if operating expenses (Opex) can be significantly reduced, then even with a decrease in gross margin, overall profitability can be maintained. The "good business" of AI no longer depends on the level of gross margin, but on whether positive cash flow and product stickiness can be maintained after scaling up. This is completely different from the traditional SaaS model that relies on licensing fees and subscription price differences. In the investor's context, low gross margins are being reinterpreted as "proof of success." In the AI ​​field, low gross margins mean that the product is indeed being used extensively. Therefore, when a company claims to be an AI company yet maintains a gross profit margin of over 80%, investors may question its actual usage. This shift in perception is also gaining consensus in the capital market: it's better to have a large scale, low gross profit margin, and rapid growth than a small-scale, high-profit business. This logic also has a workable path in corporate strategy. Traditional SaaS companies can leverage their existing high-profit businesses to subsidize the expansion of their next-generation AI product lines, thus achieving a "profit for time, time for market share" strategy. The historical transformations of Microsoft and Adobe are examples: both companies experienced a decline in gross profit margins when transitioning from on-premises licensing to cloud subscriptions, but thanks to scale and compound product growth, their stock prices more than doubled within a decade. The most common pitfall in this type of transformation is that companies overly defend their existing profit margin structure. Many companies fear being punished by the market for declining gross profit margins, ignoring the long-term valuation increases brought about by scale expansion. In the AI ​​cycle, trying to maintain the old high gross profit margin is itself a misaligned strategy. Truly smart companies view declining gross margins as a sign of business maturity, not a crisis. Some leading AI tool companies are already testing the boundaries in this way. For example, Figma announced it would promote its next-generation AI tools with lower gross margins. Investors initially questioned this, but after confirming its potential for user expansion, they responded positively. The market is learning how to reprice the relationship between growth and profit. Meanwhile, the platformization trend in AI is complicating competition at the application layer. Basic model providers (such as large cloud providers) act as both infrastructure and directly participate in applications. This significantly increases the difficulty of purely "independent application startups." Companies wanting to break through in this system must find vertical sectors not fully covered by giants—especially fragmented, non-standardized industry needs. This is also why many investors are optimistic about the small and medium-sized enterprise (SMB) services market. In this customer structure, AI tools can efficiently replace traditional manual processes, thus achieving scalable penetration even with limited profit margins per customer. For this type of market, AI is not just an automation tool, but a new form of "capability outsourcing." As industries redefine gross profit, returns, and growth, the value logic of the AI ​​industry is also being rewritten. It is no longer an extension of "software is profit," but a more complex composite model closer to the real economy. Every AI company is learning how to balance computing costs, model inference, and commercialization pace; this will be the true main theme of competition in the coming years.