The competition ultimately converged on the battle for control of three layers: the compute layer (CAPEX arms race, $805B/year), the model layer (R&D competition among Anthropic, OpenAI, and Google), and the workflow layer (the battle for entry points among Cursor, Copilot, and enterprise SaaS). The only true moat forms at the last layer— whoever controls the workflow controls the data flywheel and gains a structural advantage in the next round of model iteration. Cursor’s $60B acquisition price is essentially the market’s latest anchor for valuing “workflow control,” and this price continues to rise.

Author and source: SkillsMaster

Introduction: Three Battlegrounds and One Central Proposition

In 2026, the largest capital concentration in human history is taking place. The six major U.S. tech giants will invest $8.05 trillion (approximately RMB 5.8 trillion) in AI infrastructure this year—exceeding the annual GDP of most countries and more than double the entire U.S. defense budget for 2023.

Meanwhile, SpaceX acquired Cursor, an AI programming tool that was valued at only $590 million three years ago, in a $60 billion all-stock deal, completing the signing on the fourth day after SpaceX’s IPO, triggering a 17% single-day surge in SPCX shares and briefly surpassing Microsoft in market capitalization. Anthropic’s annualized revenue skyrocketed from $10 billion to $470 billion over 16 months, despite never having posted a quarterly profit, and its valuation has now approached $965 billion.

Behind these events are different fronts of the same war. This article breaks down the war into three distinct yet interdependent battlegrounds: the CAPEX arms race at the compute layer, the R&D competition at the model layer, and the battle for entry points at the workflow layer. The central argument is that the strength of moats varies significantly across these three layers, and most market participants are misallocating their attention to the wrong layer.

Chapter 1: The Hashrate Layer: An $805B CAPEX Arms Race

The compute layer is the material foundation of this war and the threshold for entry. In 2023, the six major U.S. hyperscalers—Amazon AWS, Google Alphabet, Microsoft Azure, Meta, Oracle Cloud, and CoreWeave—combined for $146 billion in capital expenditures; the projected figure for 2026 is $805 billion, representing a 451% increase over three years.

Jensen Huang, at GTC Taipei 2026, outlined the core economic logic of AI factories: compute equals revenue, because every token is income and every token is profit. This logic transforms CAPEX from a "cost" into a "capacity investment"—much like a factory building additional production lines. The capital expenditure for a single GW AI factory has reached $50–80 billion; NVIDIA’s Vera Rubin NVL72 rack system reduces inference costs by a factor of 10 compared to Blackwell, further reinforcing this economic model.

In 2023, these six companies spent only 40% of their operating cash flow on CAPEX, using substantial cash for stock buybacks and dividends. By 2026, this ratio exceeded 100%, meaning operating cash flow alone could no longer cover infrastructure expenditures, forcing companies to turn to external financing. Alphabet’s $84.75 billion equity financing (June 2026)—structured as a multi-tiered capital stack ($40 billion in convertible preferred shares, $10 billion in zero-coupon bonds, and $34.75 billion in common and preferred stock)—is a direct result of this pressure and the largest single equity financing in history.

The moat of the computing power layer is real, but it is an entry barrier, not a differentiating advantage. Those who possess computing power merely gain "qualification to compete," but cannot win the final competition based on this alone.

M7 stock price data (as of June 18, 2026) reveals a structural contradiction: NVIDIA is the most direct beneficiary of the CAPEX arms race, yet its stock has declined 18.9% year-to-date in 2026—the largest drop among the M7. Meanwhile, markets are pricing in a long-term risk: a significant portion of downstream buyers’ CAPEX is being allocated to building proprietary ASIC pathways that bypass NVIDIA (AWS Trainium, Google TPU v7, Microsoft Maia). The performances of CoreWeave (+240%) and Micron (+259%) year-to-date in 2026 reflect capital markets’ assessment of midstream beneficiaries in the AI supply chain.

Chapter Two: Model Layer: R&D Competition and the "Moat Illusion"

If the compute layer determines who is eligible to compete, and the model layer determines who leads in the early stages of competition—Sensor Tower’s data has already shown that the lead in the model layer cannot be converted into sustainable user lock-in.

ChatGPT's global market share declined steadily from approximately 85% in May 2023 to around 43% in May 2026, a drop of more than 40 percentage points, with no rebound at any point. This trend conveys a key signal: the network effects of consumer-facing LLMs are extremely weak. Users switch based on immediate utility, with no social lock-in such as "my friends are here, so I'm here," and no accumulated content libraries over years, like Netflix’s catalog.

In January 2025, the release of DeepSeek triggered the most dramatic single-share fluctuation in the entire time series—ChatGPT lost approximately 10 percentage points within weeks. This demonstrates that an open-source, free, and performance-equivalent alternative is sufficient to redistribute tens of millions of users in an extremely short time. The switching cost for consumer-facing LLMs is effectively zero.

Since its founding in 2021, Anthropic has never achieved a single profitable quarter, burning $56 billion in cash in 2024 alone (gross margin of -94%), yet its valuation reached $965 billion by mid-2026 (Series H). Annualized revenue grew from $1 billion in January 2025 to $47 billion in May 2026—an increase of 47 times in just 16 months.

The core of this valuation logic isn’t current profitability, but the dual lock-in created by enterprise API integrations: 80% of revenue comes from enterprise customers, and over 300,000 commercial clients have deeply embedded the Claude API into their codebases, compliance systems, and product workflows. The switching cost is no longer a question of “which model is better,” but “the engineering cost of rebuilding all integrations”—a cost that often far exceeds any performance difference between models.

The moat at the model layer is fleeting—new model releases every 6 to 12 months can erase any performance advantage. True lock-in comes from workflows and data integrations built on top of the model.

Chapter 3: Workflow Layer: The Battle for Entry Points Between Cursor, Copilot, and Enterprise SaaS

The workflow layer has the deepest moat and longest duration among the three battlegrounds. Entering the workflow means entering the environment where users spend eight hours working each day—once habits are formed, data is accumulated, and processes are embedded, the cost of replacement escalates from "whether the model is easy to use" to "rebuilding the entire work system."

In-depth Case Study | SpaceX's $60 Billion Acquisition of Cursor: From a $59 Million Seed Round to the Most Expensive AI Tool Acquisition in History

Cursor was founded in 2023 by four MIT undergraduates who forked VS Code to revolutionize the way developers interact with code through a "Vibe Coding" workflow—eliminating the need to manage low-level syntax and instead enabling high-level AI orchestration with AI assistance. At its peak, Cursor captured 41% of the AI programming tools market and accounted for approximately half of Anthropic Claude API revenue.

The most important lesson from the Cursor case is not its success, but its structural fragility. After Anthropic cut off Claude’s access in 2026, Cursor’s AI programming market share plummeted from 41% to 26%. This event clearly reveals that workflow loyalty at the application layer depends on the stability of model-layer supply; once the underlying provider reclaims control, even the strongest workflow loyalty can instantly collapse.

Acquired by SpaceX in a $60 billion all-stock transaction, this essentially addresses this supply-side risk by leveraging the balance sheet—integrating xAI’s Grok model with the Memphis Colossus supercomputer (one of the world’s largest GPU clusters) to internalize model supply, while retaining Cursor’s vast repository of real-world developer code decisions. Code generation represents the highest-value application of LLMs, and this data holds irreplaceable value for the ongoing improvement of xAI’s models.

Microsoft’s workflow strategy follows a completely different path from SpaceX/Cursor. Rather than relying on product-led organic growth, Copilot achieves forced adoption through Microsoft 365’s 345 million paid subscribers. GitHub Copilot’s annualized revenue has exceeded $2 billion (as of 2026), with enterprise renewal rates surpassing 85%.

More importantly, Microsoft’s data advantage: enterprise workflow data accumulated through products like Office, Teams, and Outlook enables a contextual understanding that no standalone AI tool can easily replicate. When Copilot can reference yesterday’s Teams meeting notes within a Word document while simultaneously linking to relevant email threads in Outlook, the switching cost escalates from “software replacement” to “disruption of work memory.”

Competition in the workflow layer extends beyond general-purpose AI tools. Traditional enterprise SaaS vendors are embedding LLM capabilities into their core products, creating verticalized AI workflow control. Salesforce Einstein GPT’s direct access to CRM data gives it far greater stickiness within sales workflows than any general-purpose LLM interface. Similarly, Workday AI’s integration into HR decision-making processes creates extremely high data and workflow migration costs.

Historical Parallel Case | WhatsApp ($22B) → Cursor ($60B): The Evolution of Network Effect Acquisitions

In 2014, Facebook acquired WhatsApp for $22 billion (comprising $4 billion in cash and $15 billion in stock), at a time when WhatsApp had a net loss of $138 million in 2013 and nearly zero revenue. The acquisition rationale: user contact lists formed a horizontal lock-in, where each new user increased the value of the entire network (a classic two-sided network effect); Facebook’s defensive motive was to prevent competitors from gaining access to the mobile messaging gateway.

Cursor’s logical structure is similar but more complex: its horizontal network effects are weaker than WhatsApp’s (developers aren’t forced to use Cursor because their colleagues do), but its vertical data flywheel is far stronger than WhatsApp’s (real code decision data continuously improves the model, creating a self-reinforcing cycle: workflow → data → model → better workflow). The $60 billion valuation reflects the market’s pricing of the combination of “workflow control + code data flywheel,” a premium of approximately 172% over WhatsApp, signaling a revaluation of workflow value in the LLM era.

Chapter 4: The AI Factory War: Location, Establishment, and Conditions for the Collapse of the Moat

Analyzing all three battlegrounds, the moat is not located in a single place. Different players have built barriers of varying strengths at different layers, but the key question is: which moat can endure across technology iteration cycles? Under what conditions will a moat fail?

Owning large GPU clusters creates a barrier to entry but does not establish a differentiated advantage—competitors can purchase the same hardware with equivalent capital. NVIDIA’s Vera Rubin platform reduces inference costs by a factor of 10, meaning the rapid decline in compute costs will further erode the value of the "more compute" moat. The moat of the compute layer will失效 under the following condition: large-scale, mature in-house ASICs (expected by 2027–2028), at which point the cost advantage of inference for hyperscale cloud providers will be significantly compressed.

Cursor’s accumulated real developer code decisions, Anthropic’s proprietary business data from enterprise API calls, and Microsoft’s enterprise workflow data gathered through Office 365 represent the most defensible assets of the AI era. The depth of the data flywheel’s moat depends on two variables: the proprietary nature of the data (whether it can be replicated or synthetically substituted by others) and the degree of coupling between the data and model improvement (whether the data genuinely drives the model’s differentiated capabilities).

Workflow control is the longest-lasting of the three moats. Its defensive logic does not rely on sustained leadership in model performance (as providers can be swapped at the model layer), but rather on the friction costs of migration—rewriting prompts, rebuilding API integrations, retraining staff, and re-passing security and compliance audits. The cumulative cost of these factors often exceeds the efficiency gains offered by a new model, creating persistent inertia and lock-in.

There are three conditions for obsolescence: ① The emergence of a disruptive workflow paradigm (e.g., a leap from "AI-assisted programming" to "AI fully autonomous programming," resetting the entire workflow logic); ② Open standardized interfaces eliminating migration costs (e.g., a unified AI Agent invocation protocol); ③ Regulatory mandates requiring data portability.

All three layers of moat are built on an implicit assumption: a stable supply chain. All seven co-designed chips on NVIDIA’s Vera Rubin platform are manufactured using TSMC’s 3nm process, and HBM4 memory comes from three Korean manufacturers. Geopolitical risks in Taiwan and export controls could disrupt the hardware supply chain at any moment—a risk that is not adequately priced into current CAPEX plans. This is the only true systemic, exogenous risk in the entire AI Factory war.

Conclusion: Who will win this war?

The core proposition of this article is systematically validated through data analysis across three chapters: the moats in AI competition are not on the same level; compute power determines eligibility to survive, lead in the model layer is fleeting, and only those who control the workflow layer can establish sustainable pricing power.

Sensor Tower data has confirmed low consumer stickiness, with ChatGPT’s market share halving over three years serving as the clearest evidence. Cursor’s $60 billion acquisition demonstrates that the market has already repriced “control over workflows,” and this valuation will continue to rise. Anthropic’s $47 billion ARR alongside ongoing losses proves that enterprise API integration creates sufficient lock-in to support valuations far exceeding current profitability.

From a competitive landscape perspective, Microsoft possesses the most balanced three-layer moat—Azure computing power, access to OpenAI models, and Office/GitHub workflow entry points; Anthropic leads in enterprise API stickiness but faces ongoing capital consumption pressures; Google’s distribution advantages (Android + Search) are difficult to replicate on the consumer side; SpaceX/xAI + Cursor’s vertically integrated approach is still in validation, but if successful, it will build the most difficult-to-unravel moat combination.

The final battle won't be about which model is smarter, but about which workflow is harder to leave. This fundamental business logic, proven in the WhatsApp era, has been magnified to a trillion-dollar scale in the LLM era.

Data Sources and Notes

1 Bank of America Analyst Team (April 2026); TrendForce Global Research (May 2026); Amazon, Alphabet, Microsoft, Meta, Oracle Q1 2026 earnings guidance. "U.S. AI Hyperscale CAPEX Sprint" chart; "Capital Squeeze: CAPEX vs. OCF" chart.

2 SpaceX/Anysphere merger announcement (June 2026); Cursor ARR official disclosure; HKU Business School Professor Chen Lin presentation materials (June 2026); Crunchbase funding database.

3 Anthropic Series H funding announcement (May 2026, valuation of $96.5 billion); HKU Business School course material "1 Trillion USD worth Loss Making Firm" data chart; Bloomberg Terminal.

4 NVIDIA GTC Taipei 2026 Jensen Huang Keynote (June 1, 2026, Taipei Music Center); NVIDIA Vera Rubin product announcement; SemiAnalysis "Vera Rubin: Extreme Co-Design" (February 2026).

5 Bloomberg Terminal; "M7 vs Micron 2026 YTD (June 18, 2026)" data chart; HKU Business School course materials; annual reports of various companies over the years.

6 Sensor Tower Global Research; Professor Chen Lin (Lin Chen), HKU Business School, presentation slides "Customer Price Sensitivity and Loyalty" (June 2026, Worldwide data).

Microsoft FY2026 earnings report; Official disclosure of GitHub Copilot ARR; Microsoft 365 paid subscriber data (Q1 2026); Satya Nadella's Investor Day presentation.

8 Facebook/Meta WhatsApp acquisition announcement (February 2014); WhatsApp 2013 financial data; HKU Business School course materials "Old Story in the Previous Cycle"; SEC-related documents.