Models, Tokens, and the Emerging Economic Framework

Recently, OpenAI's plan to go public has put the spotlight on this laboratory, which has long been operating under a 'non-profit cap' structure, thrusting it into the public market's limelight. Meanwhile, Alphabet, Google's parent company, has initiated an $80 billion financing plan, with Berkshire Hathaway alone subscribing for $10 billion.

The entry of this conservative investment giant, traditionally averse to tech stocks, into the field for the first time, signals that the AI capital game has reached a periodic high point (a phased peak). Today, it can be said that the AI industry is undergoing a profound paradigm shift.

The most evident signs are that 'funding shortages' and 'spin-offs' have become two parallel narratives for AI companies.

The former stems from computing power being a capital-intensive asset. Google's capital expenditures are projected to hit $180 billion to $190 billion in 2026, with Microsoft, Meta, and Amazon also investing hundreds of billions. Each H100 chip, transformer for a data center, and grid connection represents significant real money.

The latter has recently become a major strategy for domestic large companies. Kuaishou's Kling AI, valued at around $6 billion within the group, saw its pre-investment valuation soar to $18 billion after spinning off independently, marking a threefold increase. Baidu's Kunlunxin was spun off and listed, with external estimates suggesting it could contribute nearly $30 billion in market value increment to Baidu, equivalent to over 60% of Baidu's current total market value.

Behind these phenomena lies a redefinition of AI assets by capital. Within the consolidated financial statements of large firms, AI businesses are seen as profit-draining investments; however, once independent, they are priced based on sector scarcity, revenue growth, and future imagination space, with price-to-sales ratios of several dozen times not uncommon.

These two seemingly independent threads point to the same core: AI is shifting from a technology-driven narrative to a new competitive landscape dominated by capital efficiency.

01 The End of the Computing Power Race: The Fracture and Reconstruction of Financing Logic

Behind the 'funding crunch' lies a fundamental logical chain. Today's competition among AI large models is essentially not about product competition but a capital-intensive race in computing power scale. OpenAI has undertaken future expenditure commitments of approximately $600 billion for computing power expansion, and even after completing $122 billion in financing, this capital is expected to be exhausted within three years.

More intuitively, OpenAI's CFO Freier disclosed earlier that while annualized revenue reached over $20 billion in 2025, it is still insufficient to cover massive losses, with the company incurring about $1.22 in losses for every $1 in revenue generated.

The crux of the problem lies in the fact that the cost curve of AI businesses is fundamentally different from that of traditional internet businesses.

For WeChat, each additional user brings marginal costs close to zero; however, for ChatGPT, the more popular it becomes, the higher the inference costs due to increased usage, making user growth not only a positive but also a cost pressure. This 'anti-internet' business model means that scale effects do not bring profits but instead amplify cash flow pressures—user growth no longer directly equates to value growth.

Deeper still is the phenomenon of 'circular accounting' in the AI era: Microsoft's $13 billion investment in OpenAI was not delivered in cash but in the form of 'cloud credits,' which OpenAI uses to train models, while Microsoft counts them as new cloud revenue.

This closed-loop operation, where 'investments buy cloud services,' appears as healthy revenue growth on the surface but is essentially paying oneself with one's own money and then qualifying it as sales revenue. It is estimated that OpenAI's annual cloud service bills have swelled to over $60 billion, more than double its actual revenue of $25 billion.

This is the essential contradiction behind the 'funding crunch': the disconnect between valuation bubbles and actual cash flows. When investors begin to focus on 'free cash flow' rather than 'book profits,' the valuation system previously supported by mutual investment commitments and circular orders faces the risk of successive valuation corrections.

OpenAI plans to incur $14 billion in losses in 2026 and is not expected to become profitable until 2029, while Google's capital expenditures in 2026 are projected to reach $180 billion to $190 billion. These figures indicate that the current 'funding crunch' in AI is not merely a cyclical liquidity issue but a dilemma of an entire business model at the capital structure level.

02 Why Can One Financial Statement Be Worth Three Times as Much?

One of the most noteworthy signals in 2026 is the concentrated spin-off of core AI assets by large firms.

Kuaishou's AI video product Kling plans to conduct Pre-IPO financing at a valuation of $20 billion, nearly 70% of Kuaishou's parent company's market value. Meanwhile, Baidu is pushing its AI chip company Kunlunxin toward a dual listing on 'A+H' shares, with 2025 revenue expected to exceed 3.5 billion yuan and achieve break-even; Alibaba is reportedly planning to spin off Pingtouge, and ByteDance's Doubao may follow suit at any time.

Consider this: before the spin-off, Morgan Stanley valued Kling at only about $6 billion; after the spin-off, it is targeting $20 billion in financing. With the same assets, revenue, and team, merely changing the financial statement resulted in a valuation difference of more than three times overnight.

The change in valuation logic here reveals a structural divergence: the primary market follows a non-conventional pricing mechanism where 'consensus determines value.' The primary market looks to the future, sector status, imagination space, and whether there will be buyers in the next round, but pays little attention to current profits and revenue.

The core logic behind Kling's $20 billion valuation lies in the scarcity of such leading assets. After Sora shut down, there are only a handful of leading players left in the AI video generation sector, and the label of 'AI infrastructure for the content industry' itself commands a premium.

So, what qualifies as a leading asset currently? In today's AI landscape, it means having a self-developed foundational model (whether language, video, or multimodal) rather than a wrapper or fine-tuned version; having proven at least one vertical scenario with large-scale users or revenue (not a demo or proof of concept); and having 'exit expectations' for subsequent financing—either a strategic buyer (large firm) or an IPO pathway (US, Hong Kong, or A shares).

Companies meeting these three criteria can be counted on two hands globally. OpenAI, Anthropic, xAI, Google DeepMind (if independent), China's Zhipu, Yuezhi Dark Face, MiniMax, ByteDance's Doubao (if independent), Kuaishou's Kling (in spin-off), and Baidu's Kunlunxin (chip side). Each is a scarce target, with 'buyers lining up and sellers setting high prices.'

The underlying logic for the 'revaluation' of these companies is the cognitive shift of AI assets within large firms from 'cost centers' to 'value centers.'

Within large firms, AI businesses are seen as part of the group's operations and are typically classified as 'strategic investments,' meaning their costs (computing power, R&D, data annotation) are lumped together with the group's mature cash flow businesses (such as advertising, e-commerce, gaming). The group's CFO looks at the consolidated financial statements, and as long as the AI business is burning money, it is constantly asked to explain 'when it will contribute net profits.'

In this context, AI teams are forced to make short-term ROI arguments, and their valuation logic is naturally suppressed under the shadow of the group's overall PE multiple—mature internet companies typically only receive 10-15 times PE. Even high-growth businesses can only enjoy a 20% premium, not the 3-5 times PS of independent sectors.

Once spun off independently, the independent financial statements can redefine the boundaries of 'costs' and 'revenues.' For example, computing power costs previously incurred internally can now be re-priced as 'related-party transaction revenue' at market prices; model training previously expensed as R&D can now be capitalized as 'intangible assets' and amortized over time.

In other words, these assets acquire the pricing model of 'growth companies.' The spun-off AI companies can more flexibly raise funds and advance strategies, avoiding the constraints of internal resource allocation within the group and obtaining independent pricing in the capital market based on their own growth prospects.

At the same time, this involves further differentiation in the valuation system. The existing businesses of large firms, superimposed with AI labels, begin to exhibit new potential for premium valuation in the secondary market.

This also explains why traditional internet giants (such as Baidu at $47.5 billion and Kuaishou at $27 billion) are being caught up or even surpassed in market value by AI newcomers—Zhipu's latest market value is approximately $58.6 billion, surpassing Baidu to become China's ninth-largest AI tech stock.

03 From 'Model Worship' to 'Value Realization': A Structural Shift in Industry Narrative

Some experts have noted that the rapid development of the current AI era is similar to the previous mobile internet boom, a precise analogy, but the key difference lies in the fundamental distinction in cost structures.

The mobile internet boom relied on the proliferation of smartphones and declining bandwidth costs, with downward marginal costs; the AI boom faces hard constraints such as rising computing power costs, surging power consumption, and lengthy data center construction cycles.

One observation is that the current AI industry is in a state of 'water at 85 degrees Celsius, about to boil but not yet boiling.'

The directions for technological breakthroughs (agents, multimodality) are clear, and investments in computing power infrastructure are unprecedented, with capital expenditures by large US hyper-scale cloud computing companies reaching $805 billion in 2026, nearly doubling from the previous year's forecast. However, true commercialization and scale deployment are still at a critical point of imminent breakthrough.

Currently, only a small percentage of CFOs saw actual financial value from AI in 2025, and even fewer Chinese companies have achieved revenue growth through AI. This tension of 'high investment, low return' is precisely a signal of the industry's transition from hype to practical implementation.

Many may not have noticed that the weight of the AI value chain has shifted from the GPU side to the entire system side. Morgan Stanley's latest research points out that 'agent AI marks a structural shift from computation to orchestration,' with CPU-side orchestration time accounting for 50% to 90% of total latency in agent workflows, projecting an additional $32.5 billion to $60 billion in CPU market space by 2030.

This means that the core contradiction of the industry is shifting from 'insufficient computing power' to 'insufficient system efficiency,' and the corresponding investment logic will expand from 'single-chip computing power race' to 'full-stack system engineering,' with GPUs determining 'what can be done,' but CPUs and systems determining 'what can be profitable.'

If the mobile internet boom was driven by 'connection,' the AI boom will be driven by 'intelligence,' with a value chain breadth likely exceeding that of the mobile internet, covering computing power, models, applications, data, and the entire chain.

Some economists have pointed out that 2026 is becoming a critical inflection point for AI to transition from 'assisted thinking' to 'autonomous execution.' The current core contradiction is shifting from 'who can train the strongest model' to 'who can be the first to convert AI capabilities into deployable commercial value and user benefits in the most cost-effective, fastest, and broadest manner.'

'It's not just about redefinition but also revaluation.' Everything happening in the AI industry in 2026—giants facing funding shortages, frantic financing, large firms spinning off assets, and a rush of IPOs—is essentially a concentrated release of the same capital logic: when the path of 'burning money for growth' reaches its end, the industry must answer the most fundamental question: How much is this technology really worth?

The answer to this question will determine the power dynamics of the AI industry for the next decade. And 2026 is precisely the moment when this competition between capital and technology unfolds in full swing.

This article is an original work by Xinmou. For reprint authorization and business cooperation, please contact us.

— END —