Overview

The competitive landscape of AI infrastructure has undergone a fundamental transformation in late 2025 and the first half of 2026. Google has emerged not merely as a customer of Nvidia's silicon but as a direct competitor across hardware, cloud services, and software platforms. The strategic moves announced between Google Cloud Next 2026 (April 22) and Google I/O 2026 (May 19) represent a coordinated assault on Nvidia's dominance in AI accelerators, CoreWeave's position in the GPU cloud market, and Nebius's ambitions in European and specialized AI cloud services. This report provides a detailed analysis of Google's specific announcements, their implications for each competitor's business model, and the broader market dynamics that will shape the AI infrastructure industry through 2027 and beyond.

Google's Strategic Announcements and Infrastructure Buildout

Eighth-Generation TPU Architecture: TPU 8t and TPU 8i

At Google Cloud Next 2026 on April 22, Google unveiled two distinct eighth-generation Tensor Processing Units: the TPU 8t designed specifically for training workloads and the TPU 8i optimized for inference [1][2]. This bifurcation of training and inference into separate silicon represents a strategic departure from Nvidia's general-purpose GPU approach and from Google's own previous TPU generations, which were unified architectures. The chips are purpose-built for matrix-heavy workloads and agentic AI applications, offering superior efficiency for a narrower set of tasks compared to Nvidia's broadly applicable GPUs [3].

The rapid iteration cadence is itself a competitive weapon. PitchBook reported that Google has rolled out four TPU generations in roughly three years, compressing the useful life of any given chip and making it difficult for competitors to ascertain residual value or plan long-term procurement strategies [4]. Google's sixth-generation TPU, codenamed Trillium, continues to power internal workloads and has been considered for experimental orbital AI compute deployment under Project Suncatcher, a partnership with SpaceX to place solar-powered Trillium TPUs in low Earth orbit [5].

TPU Commercialization: Selling Chips Externally

In a landmark decision announced during Alphabet's Q1 2026 earnings call on April 29, CEO Sundar Pichai confirmed that Google will begin delivering its custom TPU chips to a select group of customers for deployment in their own data centers, with initial shipments expected in the second half of 2026 and the majority of revenues anticipated in 2027 [6][7]. Morgan Stanley estimated that selling 500,000 TPU chips could add approximately $13 billion in Google revenue by 2027 [6]. Citizens analyst estimates suggest Alphabet's TPU business could generate $3 billion in 2026 revenue, rising to $25 billion in 2027 [8].

Pichai framed the decision in terms of economies of scale: "Some of it helps us get more economies of scale, scale in our overall compute environment as well. And so, it helps us invest in the cutting edge, which we need to do with the next generation" [7][9]. The move mirrors Amazon's trajectory with Trainium chips; Amazon CEO Andy Jassy stated in his April 2026 shareholder letter that "virtually all AI thus far has been done on Nvidia chips, but a new shift has started" and confirmed that Amazon may begin offering full racks of Trainium chips outside its cloud "over the next couple of years" [6][10].

The Blackstone-Google Joint Venture

On May 19, 2026, Blackstone, the world's largest alternative asset manager with over $1.3 trillion in assets under management, announced a joint venture with Google to create a new U.S.-based AI infrastructure company [11][12][13]. The terms are striking in their ambition:

• Blackstone will commit an initial $5 billion in equity capital through its dedicated AI investment arm, Blackstone N1 [11][12]
• Google will supply its proprietary TPUs, hardware, software, and operational expertise [11][12]
• The first 500 megawatts of compute capacity is expected online by 2027, with plans to scale significantly [11][12]
• The Wall Street Journal reported that the venture could support approximately $25 billion of total investment including leverage [4][14]
• The unnamed company will be led by Benjamin Treynor Sloss, a Google executive with over 20 years of experience building and operating Google's global infrastructure [11][12]
• Blackstone holds a majority stake [11]

Jon Gray, Blackstone's President and COO, described the opportunity in generational terms: "We see a generational opportunity to invest capital at scale building AI infrastructure. This new company has enormous potential as it helps to meet the unprecedented demand for compute" [11][12]. Thomas Kurian, CEO of Google Cloud, emphasized the TPU value proposition: "This joint venture with Blackstone helps meet growing demand for TPUs, which are optimised specifically for efficiency and performance in the AI era" [12][13].

The venture has been explicitly framed as a competitive challenge to Nvidia, CoreWeave, and Nebius [12][15]. The joint venture will offer compute-as-a-service infrastructure combining data center operations, networking, and TPU access outside the traditional Google Cloud environment [12][13][15]. Rittenhouse Research characterized the deal as a "strong endorsement of the neocloud model" itself, while Rockefeller's Jason Kotik offered a more ominous assessment: "Welcome to Hell... Google and Blackstone joining forces means every other AI company just got put on notice" [16][17].

Google's $40 Billion Anthropic Investment

On April 24, 2026, Google announced plans to invest up to $40 billion in AI firm Anthropic, comprising an immediate $10 billion at a $350 billion valuation, with an additional $30 billion contingent on meeting performance targets [18][19]. The deal includes Google Cloud providing 5 gigawatts of compute capacity over five years, building on a prior partnership with Broadcom for 3.5 gigawatts of TPU-based capacity starting in 2027 [18][19]. This single deal nearly doubled Alphabet's cloud backlog to $462 billion, though analysts have cautioned about concentration risk—Gil Luria of D.A. Davidson noted that Alphabet "did it the same way Oracle did" in reporting a doubled backlog without disclosing that nearly the entire increase came from one customer [20][21].

Anthropic also secured a separate $5 billion investment from Amazon as part of a broader agreement under which Anthropic is expected to spend up to $100 billion for approximately 5 gigawatts of compute capacity over time [18][19]. The company is reportedly considering an IPO as soon as October 2026 [18][19].

Direct TSMC Relationship and Supply Chain Independence

On May 21, 2026, it was reported that Alphabet is building direct relationships with Taiwan Semiconductor Manufacturing Company to design custom AI chips independently, mirroring Apple's approach to chip design and manufacturing [22]. This strategic shift reduces Google's dependence on third-party chip designers and potentially positions it to secure more favorable manufacturing allocations from TSMC as TPU demand grows [4][22].

Software Stack Transformation: Gemini Enterprise Agent Platform and JAX

Google made a significant architectural declaration at Cloud Next 2026 by announcing the Gemini Enterprise Agent Platform, simultaneously retiring Vertex AI as a standalone brand [23][24]. The platform is built on four pillars—Build, Scale, Govern, and Optimize—and positions AI agents as the primary unit of work in enterprise technology stacks [23][24]. Key elements include:

• Agent Studio for low-code visual agent construction
• Agent Development Kit (ADK) supporting Python, Java, and Go, processing over six trillion tokens monthly
• Agent Identity, which assigns every agent a unique cryptographic credential based on the SPIFFE standard (X.509 certificates), creating verifiable audit trails
• Agent Gateway enforcing policies against prompt injection and data leakage [23][24]

Forbes noted that "the lock-in sits at the runtime and governance layer, not at the model layer where Google explicitly supports third-party models" [23][24]. This is a sophisticated competitive strategy: by making the governance and orchestration layer sticky while supporting over 200 models including Anthropic's and third-party offerings, Google can attract multi-model customers while building dependency on its platform.

At Google I/O 2026 on May 19, Google deepened its software offensive with Antigravity 2.0, an updated agentic coding application with a new desktop app, CLI tool, SDK for custom workflows, and native voice command support—all powered by the Gemini 3.5 Flash model [25]. Gemini Spark, an always-on AI agent running 24/7 on virtual machines, was introduced as a persistent workforce tool [26][27]. The Google Search overhaul was described as the biggest upgrade to the search box in 25 years, reimagined by AI with generative UI, information agents, and mini apps [26][28].

JAX continues to power Google's AI infrastructure, providing superior scaling efficiency for large model training on TPU hardware. Together with the Pathways orchestration system, Google's software stack offers tight integration between hardware and software that GPU-only providers cannot replicate without Nvidia's CUDA ecosystem.

Alphabet Q1 2026 Earnings and Infrastructure Spending

Alphabet reported stellar first-quarter earnings on April 29, 2026, significantly exceeding analyst expectations [29][30][31]:

• Earnings Per Share: $5.11 versus consensus of $2.63, a 94.1% beat [29][30][31]
• Revenue: $109.9 billion, up 22% year-over-year [29][30][31]
• Google Cloud Revenue: $20.03 billion, up 63% year-over-year, beating expectations of $18.05-18.4 billion [29][30][31]
• Capital Expenditures (Q1 2026): $35.67 billion, more than doubled year-over-year [30][32]
• Cloud Backlog: Doubled to $462 billion [20][21][30]
• AI Token Throughput: 16 billion tokens per minute via the API [30]

Pichai stated explicitly that Google Cloud was "compute constrained in the near-term" and that "our cloud revenue would have been higher if we were able to meet that demand" [30][33]. The company raised its full-year 2026 capital expenditure estimate to between $175 billion and $185 billion, potentially double 2025 levels [30][32][34]. The combined 2026 capex commitment across the five major hyperscalers is now on track to exceed $650 billion [4][30][34][35].

Alphabet's market capitalization surged past $4.6 trillion, and the gap with Nvidia narrowed to roughly $200 billion [29]. Options traders assigned a 53% probability that Alphabet would overtake Nvidia as the world's most valuable company [29]. The Next Web concluded: "Nvidia built the engine. Alphabet is building the car, the road, and the toll booth. The market is starting to price accordingly" [29].

Impact on Nvidia

Direct Hardware Competition: TPU Commercialization and Customer Defection Risk

The most significant competitive threat to Nvidia from Google's recent moves is the commercialization of TPU chips for external deployment. Sundar Pichai's announcement that Google will deliver TPUs to select customers for use in their own data centers represents a fundamental shift from Google as a consumer of Nvidia's chips to a direct competitor in the silicon market [6][7]. Citizens analyst estimates project Alphabet's TPU business could reach $25 billion in 2027 revenue, which would represent a meaningful displacement of what would otherwise be Nvidia GPU sales [8].

The Blackstone-Google joint venture compounds this threat by creating a new channel for TPU-based compute that competes directly with Nvidia's GPU cloud business and, by extension, Nvidia's hardware sales to cloud providers [11][12]. The venture will offer TPUs as a service outside the traditional Google Cloud environment, potentially capturing demand that would otherwise flow to Nvidia-powered infrastructure [12][13][15].

Amazon's parallel development reinforces the industry trend. Amazon CEO Andy Jassy stated there is "a good chance" Amazon will start offering full racks of Trainium chips outside its cloud "over the next couple of years" [6][10]. Amazon's chips business has crossed a $20 billion annual revenue run rate, growing at triple-digit pace year-over-year, and Jassy stated that if the chips business were a stand-alone company, its revenue run rate would be $50 billion [36]. Meta is also deploying homegrown chips, and Broadcom has disclosed long-term agreements to design custom accelerators for both Google (through 2031) and OpenAI (10-gigawatt partnership) [8][37].

The structural shift is driven by the AI industry's move toward inference workloads, where Google pitches its TPUs as more cost-effective than Nvidia GPUs [6][38]. Forrester analyst Alvin Nguyen noted: "Selling products is very different than access to them," emphasizing that Amazon and Google would need to provide education, support, and enterprise-grade service for external chip buyers [6]. However, senior analyst Beatriz Valle of GlobalData concluded: "This process will take years but it is irreversible now" [6][38].

Nvidia's Financial Position and Counter-Arguments

Nvidia's recent financial results demonstrate that, for now, demand remains insatiable. On May 20, 2026, Nvidia reported record quarterly revenue of $81.6 billion for the period ending April 26, 2026, up 85% year-over-year [39][40][41]. Data center revenue reached $75.2 billion, up 92% year-over-year [39][40][41]. Adjusted EPS rose 140% year-over-year, gross margin stood at 75%, and free cash flow hit a record $48.6 billion [39][41]. The company guided Q2 revenue of approximately $91 billion, representing approximately 95% year-over-year growth [39][41]. CEO Jensen Huang described demand as "parabolic" and stated: "The buildout of AI factories—the largest infrastructure expansion in human history—is accelerating at extraordinary speed" [39][41][42].

Nvidia's competitive moat remains formidable. The CUDA ecosystem is deeply embedded across enterprise AI workloads, creating high switching costs for customers [43]. Nvidia's networking business has grown to a $60 billion annualized run rate, driven by Q1 networking revenue of $15 billion (tripling year-over-year), positioning it to overtake Broadcom, Cisco, and Arista in AI networking [44]. CEO Huang stated on the earnings call: "We don't let noise around competition distract us" [42]. Nvidia's next-generation Vera Rubin platform promises 5x inference performance and 10x lower cost per token compared to Blackwell, potentially reasserting Nvidia's price-performance leadership [45][46].

However, forward-looking concerns are mounting. Despite consistently beating earnings expectations, Nvidia's stock has declined after each of the last four quarterly reports, including a 1.5-2% drop following the May 20 beat [39][41]. Bank of America analyst Vivek Arya noted that most major customers "appear to be putting equal emphasis on heterogenous deployments of both Nvidia's chips and custom-made chips" [8]. Clayton Allison of Prime Capital Financial observed: "I wouldn't say that Nvidia's competitive positioning is materially threatened by these new chips, but the market action in Nvidia reflects how people are starting to question its market share, its competitive moat, and its margins" [8].

Pricing Power Assessment

There is no evidence of immediate pricing pressure on Nvidia's current-generation GPU products. DigitalOcean CEO Paddy Srinivasan reported in May 2026 that "NVIDIA H100/H200 prices are still appreciating" [47]. CoreWeave CEO Mike Intrator similarly stated that his company is "sold out in H100s" and "sold out in A100s" and "seeing price appreciation as more inference is coming in" [48]. Nvidia's gross margin remained at 75% in Q1 FY2027, consistent with historical levels [39][41].

The risk is forward-looking. Nvidia's gross margin guidance of "mid-70% margins for the year" includes potential pressure in the second half as the Vera Rubin platform rolls out, though this relates to product transition dynamics rather than direct competition from custom chips [49]. The Motley Fool noted that "pricing power could soften as alternatives become more credible—and at Nvidia's current valuation, the stock doesn't seem to have much room for margin compression" [36]. Nvidia's forward P/E of approximately 45 leaves limited room for error [39].

Customer Concentration Risk

Nvidia disclosed in its filings that two customers accounted for 36% of FY26 revenue, highlighting concentration risk [43]. These customers—widely believed to be Microsoft and Google, or possibly Amazon—are precisely the hyperscalers most aggressively developing custom alternatives. Any slowdown in AI spending by these major customers, or any shift in their internal procurement toward their own chips, could pressure Nvidia's revenue growth [43].

The irony is striking: Nvidia's largest customers are now its most credible competitors. This dynamic is without modern precedent in the semiconductor industry. As Bill Stone of Glenview Trust Company put it: "The problem with having basically 100% market share is that there's only one direction for it to go" [8].

Impact on CoreWeave

Business Model Vulnerability: Nvidia Dependency

CoreWeave's business model is built on a straightforward proposition: provide access to Nvidia GPUs with superior performance, flexibility, and pricing compared to hyperscaler clouds. This model is now under attack from multiple directions. The most direct threat comes from the Google-Blackstone joint venture, which created a TPU-based compute-as-a-service offering explicitly positioned as a competitive challenge to CoreWeave and other neocloud providers [12][15].

CoreWeave's dependence on Nvidia is both a strength and a critical vulnerability. The company is an "NVIDIA Exemplar Cloud" for inference on the GB200 N72 platform and received a $2 billion equity investment from Nvidia in early 2026 [48][50]. The partnership commits CoreWeave to building over 5 GW of AI factories by 2030 [48][50]. This deep integration provides privileged access to Nvidia hardware but leaves CoreWeave unable to offer customers an alternative to Nvidia's ecosystem. If Google's TPU offering captures a meaningful share of the AI inference market—where Google claims superior cost efficiency—CoreWeave has no architectural alternative to offer its customers.

Financial Health and Margin Trajectory

CoreWeave's Q1 2026 earnings, reported in May 2026, reveal a company growing rapidly but under significant financial strain [48]:

Metric	Q1 2026	Q1 2025	Change
Revenue	$2.08B	$982M	+112%
Net Loss	($740M)	($315M)	Widened
Adjusted EBITDA	$1.157B (56% margin)	$606M (62% margin)	Margin compressed 6pp
Total Debt	~$25B	~$12B (est.)	Doubled
Revenue Backlog	$99.4B	$66.8B (Dec 2025)	+49%

The adjusted EBITDA margin decline from 62% to 56% is a critical signal. CFO Nitin Agrawal characterized the dynamic as "timing-based, not economic," arguing that margins are temporarily depressed due to upfront build-out costs [48]. However, the company's debt-to-EBITDA ratio of 8.87 with aggregate debt between $20-30B creates significant financial vulnerability [51][52]. Total debt of $25B at SOFR + 2.25% to approximately 5.9% fixed rates generates substantial interest expense that depresses net income [48][51].

The company's Q2 2026 guidance of $2.45-2.6 billion fell below Wall Street's consensus of approximately $2.7 billion, causing the stock to drop 11% [48][53]. Adjusted EPS of negative $1.12 missed expectations of negative $0.91 [48][53]. Bank of America expects operating margin to improve sequentially throughout 2026, reaching 8% for the full fiscal year, but this remains far below the levels needed to sustain long-term debt service [48][54].

Pricing Competition from Google Cloud

Contrary to expectations of immediate pricing pressure, CoreWeave is not currently being undercut by Google on price. Google Cloud CEO Sundar Pichai stated explicitly that the cloud is "compute constrained" and that "revenue would have been higher if we were able to meet that demand" [30][33]. This capacity constraint means Google cannot currently serve all available demand, protecting pricing for neocloud providers.

The threat becomes acute when Google's capacity expansion comes online. The Blackstone-Google JV aims to bring 500 MW of TPU capacity online by 2027, and Alphabet's total 2026 capex of $175-185 billion dwarfs the entire neocloud sector's combined investment capacity. When this capacity comes online, Google has both the incentive and the balance sheet to compete aggressively on price. The Motley Fool noted that "the very customers funding this surge are increasingly building or commissioning their own artificial intelligence (AI) chip alternatives" [36].

Customer Concentration Risk

CoreWeave faces severe customer concentration risk. OpenAI represents approximately one-third of CoreWeave's contracted revenue, and 67% of 2025 revenue came from Microsoft [48][51][55]. This concentration creates existential vulnerability: if either Microsoft or OpenAI shifts a meaningful portion of their compute demand to internal infrastructure or to a competitor like the Google-Blackstone JV, CoreWeave's revenue could be severely impacted.

The Microsoft relationship is particularly complex. Microsoft is simultaneously a hyperscaler competitor, a major investor in OpenAI (which competes with Google's Gemini and Anthropic's Claude), and CoreWeave's largest customer. Microsoft is also building its own AI infrastructure at enormous scale. If Microsoft determines that owning its AI compute infrastructure is strategically preferable to renting from CoreWeave, the neocloud could lose its anchor customer.

Software Stack and Performance Competition

Google's software stack advantages—JAX for high-performance numerical computing, Pathways for distributed computation, and the Gemini Enterprise Agent Platform for AI agent orchestration—create a performance and developer experience gap that CoreWeave cannot easily bridge. CoreWeave offers Nvidia GPUs with standard cloud infrastructure, but it cannot offer the tight hardware-software integration that Google achieves with its vertically integrated TPU-JAX-Pathways stack.

The F5 2026 State of Application Strategy report found that 86% of organizations distribute AI workloads across on-premises, public cloud, and colocation, and 52% chain multiple AI models together [56]. This multi-cloud reality creates opportunities for Google to capture workloads that benefit from TPU-optimized software without requiring customers to abandon GPU deployments entirely.

Existential Risk Assessment

Former Deloitte cloud strategy chief David Linthicum offered a sobering assessment of neocloud viability: "If I was running CoreWeave... my concern would be the runway runs out... lenders want their money back before I'm able to hit profitability, and then I have to sell the company on the cheap, and it becomes a division of Amazon" [57]. This scenario—a distress sale to a hyperscaler—represents a plausible outcome if demand plateaus or pricing compresses before CoreWeave achieves sustainable profitability.

McKinsey has characterized neocloud bare-metal-as-a-service economics as "fragile," and the debt-laden model creates a narrow path to sustainable profitability [58][59]. Synergy Research's John Dinsdale celebrated the sector's growth in Q1 2026, but the sheer scale asymmetry between hyperscalers ($710 billion combined capex in 2026) and neoclouds (less than $50 billion in announced capacity) makes the long-term competitive outlook challenging [60].

Impact on Nebius

European Market Positioning Under Pressure

Nebius (formerly Yandex Cloud) has positioned itself as a European AI cloud provider with distinct advantages: own data center design, strategic European locations, and partnerships with major technology companies. The company operates a 310 MW AI data center in Finland, one of Europe's largest, and has secured a $2.6 billion deal with Bloom Energy for fuel-cell data center power, addressing the power constraints and high European electricity prices that are critical infrastructure challenges [61][62][63].

The Wall Street Journal noted in May 2026 that Nebius "benefits from designing data centers and the equipment inside from the ground up," providing cost and efficiency advantages over providers that rely on standard colocation facilities [61]. The company has also secured a $2 billion investment from Nvidia and a $27 billion infrastructure deal with Meta, demonstrating that major technology companies view Nebius as a credible partner for large-scale AI compute [64][65].

However, Google's European cloud expansion creates direct competitive pressure. Google Cloud grew 63% overall in Q1 2026, and the company is expanding European data center capacity, including AI-optimized regions. Non-European hyperscalers currently control approximately 70% of Europe's cloud market, according to a GITEX AI EUROPE study from May 2026 [66]. Europe's ICT market is valued at €1.02 trillion, and Germany alone may need to triple data center capacity by 2030, requiring €60 billion in investment [66][67]. The European Union's €200 billion InvestAI programme funds five AI gigafactories with 100,000+ GPUs each, creating both opportunity and competition for Nebius [68].

Sovereign AI Opportunity and Threat

The European desire for sovereign AI infrastructure—cloud services that comply with European data regulations and are not controlled by U.S.-based hyperscalers—creates a potential moat for Nebius. European enterprises and governments may prefer a European-owned cloud provider for sensitive AI workloads, particularly in sectors like defense, healthcare, and finance.

However, this sovereign advantage cuts both ways. Google has invested heavily in European data center compliance and offers sovereign cloud solutions. Google's scale ($185 billion capex vs. Nebius's limited capacity) creates an enormous resource asymmetry that could allow Google to outspend Nebius on regulatory compliance, data center construction, and price competition. The EU's digital sovereignty initiatives might ultimately benefit Nebius if regulators mandate European ownership for certain workloads, but the window for establishing a defensible position is narrow.

Financial Health and Debt Burden

Nebius's financial position is less transparent than CoreWeave's, but the available data reveals significant strain. The company issued $4.34 billion in debt in March 2026, causing a 20% stock decline [69][70]. The stock traded at approximately $147.16 following the debt issuance, with Citi maintaining a target price of $169 [69]. Building a 1+ GW AI factory in Missouri alongside the existing 310 MW facility in Finland requires enormous capital investment that the debt issuance only partially funds.

The company's reliance on Nvidia for GPU supply and on Nvidia's ecosystem for software creates the same architectural vulnerability that CoreWeave faces. Nebius cannot offer customers an alternative to Nvidia GPUs, while Google can offer both TPUs and GPUs on Google Cloud and now through the Blackstone JV.

Google's Dual Chip Strategy vs. GPU-Only Positioning

Google's ability to offer both TPUs and Nvidia GPUs on Google Cloud, combined with TPU access through the Blackstone JV, creates a strategic advantage that GPU-only providers like Nebius cannot match [12][13]. Google touts its TPUs as purpose-built for efficient processing of narrower AI applications, particularly agentic AI, with better performance-per-dollar for certain workloads [12][15]. The TPU 8i inference chip is specifically designed to compete on cost efficiency for running trained models, which is the fastest-growing segment of the AI compute market.

For customers optimizing for total cost of ownership, the ability to choose between TPU-based and GPU-based infrastructure based on workload characteristics is a significant advantage. Nebius can only offer GPUs, and while GPUs remain the dominant architecture for AI training, the shift toward inference and the growing maturity of the TPU ecosystem could progressively erode Nebius's addressable market.

Partnership Strategy as Defense

Nebius has responded to competitive pressure by deepening partnerships with Nvidia and major AI customers. The $2 billion Nvidia investment aligns Nvidia's incentives with Nebius's success, ensuring continued access to GPU supply [64]. The $27 billion Meta deal provides a long-term demand anchor that partially insulates Nebius from market volatility [65].

The Bloom Energy partnership addresses a critical constraint for European data center operations: power availability and cost [62][63]. European electricity prices are significantly higher than in the United States, and fuel-cell technology offers a path to stable, potentially lower-cost power that could provide a competitive advantage over hyperscalers importing power from the grid.

However, these defensive measures do not address the fundamental asymmetry in scale. Alphabet's 2026 capex of $175-185 billion is approximately 40 times larger than the capital Nebius could reasonably deploy. When Google decides to compete aggressively on price for European AI cloud workloads, Nebius's partnership strategy provides limited insulation.

Market Dynamics and Pricing Pressure

AI Chip Market: GPU Dominance Under Structural Challenge

Nvidia commanded an 86% share of the AI accelerator market in 2025, unchanged from 2024 [8]. However, the trajectory is shifting. Multiple credible analysts and executives have stated that the shift toward custom silicon is irreversible, even if it takes years to play out [6][8][38]. The qualitative evidence of structural change is accumulating rapidly:

• Google is commercializing TPUs externally, with Morgan Stanley projecting $13 billion in additional revenue from selling 500,000 chips by 2027 [6]
• Amazon has crossed a $20 billion annual revenue run rate for its chips business, with Trainium3 shipping and Trainium4 largely reserved [10][36]
• Meta is deploying homegrown chips and signed a deal for millions of Amazon Graviton CPUs [8][71]
• Broadcom has long-term design agreements with Google (through 2031) and OpenAI (10-gigawatt partnership) [37]
• Microsoft is developing its own AI accelerators, though details remain limited [32]

Nvidia's revenue growth is projected at 70% in its current fiscal year, slowing to 32% in fiscal 2028 [8]. The deceleration reflects the arithmetic of an enormous base ($91 billion quarterly revenue in Q2 guidance) and the gradual impact of custom silicon displacement. CEO Huang updated the outlook to more than $1 trillion in combined Blackwell and Rubin revenue through calendar year 2027, up from approximately $500 billion for 2025 and 2026 combined [49].

The inference shift is the critical catalyst for custom silicon adoption. Inference workloads are typically less demanding than training, making them more addressable by custom chips [6][38]. As AI moves from training foundation models to running inference at massive scale for millions of users, the addressable market for TPUs, Trainium, and other custom accelerators expands dramatically.

AI Cloud Infrastructure Market: Hyperscaler Dominance Intensifying

Total cloud infrastructure spending reached $129 billion in Q1 2026, with neoclouds holding approximately 5% of the market according to Synergy Research [60][72]. The growth rates tell a compelling story:

Provider	Revenue (Q1 2026)	Growth YoY
AWS	$37.6B	+28%
Microsoft Azure	~$36B (est.)	+40%
Google Cloud	$20.03B	+63%
CoreWeave	$2.08B	+112%
All Neoclouds (combined)	~$6.5B	~100%+

The neocloud sector is growing faster than the hyperscalers, but from a tiny base. The absolute revenue gap is enormous and widening in dollar terms. Google Cloud alone added approximately $7.7 billion in revenue year-over-year in Q1 2026, which exceeds the total revenue of the entire neocloud sector [29][30][31].

Wolfe Research noted in April 2026 that "public markets' eyes are moving to AI beneficiaries, and neoclouds are becoming increasingly relevant" [73]. However, the resource asymmetry is overwhelming. Combined 2026 AI infrastructure commitments across the four largest hyperscalers total approximately $710-725 billion [34][35][74]. The total capital that neoclouds could deploy in 2026 is unlikely to exceed $50 billion, and much of that is debt-financed.

Pricing Pressure Assessment

The current pricing environment does not show evidence of compression from Google's initiatives. GPU prices for Nvidia H100 and H200 are appreciating, CoreWeave is sold out across its product line, and Google is capacity-constrained [30][47][48]. The market is characterized by insufficient supply relative to demand, which protects pricing for all participants.

The risk is forward-looking and centers on three dynamics:

1. Capacity coming online: When the massive hyperscaler capex investments translate into operational compute capacity, supply will increase substantially. PitchBook declared that the "enterprise AI super-cycle is officially here" and that hyperscaler capex is heading toward $1 trillion by 2028 [4]. When this capacity arrives, hyperscalers have the incentive to fill it, potentially through aggressive pricing.

2. Custom silicon cost advantages: Google's TPUs are 2-3x more energy efficient than GPUs for AI workloads, according to semiconductor supplier LoveChip [3][75]. If this efficiency translates into significantly lower pricing for inference workloads, Google could undercut GPU-based providers while maintaining healthy margins.

3. Hyperscaler willingness to compete on price: Hyperscalers can afford to operate cloud infrastructure at lower margins than neoclouds because they monetize the compute through multiple layers—advertising (Google), e-commerce (Amazon), software subscriptions (Microsoft). A neocloud like CoreWeave or Nebius has a single revenue source: renting compute. A 10% price cut by Google might be an acceptable competitive tactic; the same cut at CoreWeave could be existential.

Margin Trajectory for CoreWeave and Nebius

CoreWeave's adjusted EBITDA margin declined from 62% to 56% year-over-year as technology and infrastructure costs jumped 127% [48]. CFO Nitin Agrawal's characterization of this as "timing-based" may be accurate in the short term, but the structural dynamics suggest ongoing margin pressure:

• Debt service costs: $25 billion in debt at current interest rates generates approximately $1.5-2 billion in annual interest expense, creating a fixed cost burden that limits margin flexibility [48][51]
• Customer concentration: The loss of any major customer would create severe revenue disruption and likely compress margins as CoreWeave seeks replacement demand [48][55]
• Commoditization risk: As GPU supply becomes more abundant and hyperscalers expand capacity, GPU compute may become more commoditized, reducing pricing power

Bank of America expects CoreWeave's operating margin to improve to 8% for FY2026 [54]. For context, AWS operates at approximately 30%+ operating margins, and Google Cloud is approaching profitability at scale. An 8% operating margin provides limited cushion against competitive pressure.

For Nebius, the margin picture is less clear but likely similar. The $4.34 billion debt issuance in March 2026, combined with the capital requirements of building gigawatt-scale data centers, suggests the company faces comparable margin dynamics to CoreWeave [69][70].

Market Share Projections: Methodological Challenges

Projecting market share shifts in the AI chip and cloud markets requires acknowledging significant uncertainty. Several credible analysts have offered conflicting assessments:

• Bearish for Nvidia: Bill Stone (Glenview Trust) expects market share erosion but not a fatal blow [8]. Beatriz Valle (GlobalData) considers the shift irreversible but long-term [6][38].
• Bullish for Nvidia: Christopher Rolland (Susquehanna) raised Nvidia's price target to $275, projecting 21% upside, and noted that CEO Huang updated the outlook to more than $1 trillion in combined Blackwell and Rubin revenue through CY2027 [49].
• Cautionary: Clayton Allison (Prime Capital) noted that market action reflects questioning of Nvidia's moat and margins [8]. Alvin Nguyen (Forrester) stated Nvidia should be "concerned but not worried" [6].

The most reasonable synthesis is that custom silicon will progressively erode Nvidia's market share from approximately 86% to perhaps 65-75% over the next three to five years, but the total addressable market is growing so rapidly that Nvidia's absolute revenue may continue to grow even as its share declines. This creates a complex investment thesis: Nvidia may be a faster-growing revenue story but a declining-margin and declining-share story simultaneously.

For neoclouds, the risk is more existential. The sector may grow from 5% to 10-15% of the cloud market as AI compute demand expands, but the debt-laden financial models of current leaders leave them vulnerable to acquisition by hyperscalers or restructuring if demand plateaus.

Conclusion

Google's strategic moves in late 2025 and early 2026 represent a coherent, multi-pronged assault on the existing AI infrastructure order. The TPU 8t and TPU 8i chips challenge Nvidia on architectural efficiency and cost of inference. The external commercialization of TPUs attacks Nvidia's hardware business model. The Blackstone-Google joint venture creates a new competitive front against CoreWeave and Nebius by offering TPU-based compute-as-a-service. The $40 billion Anthropic investment secures a major AI customer for the TPU ecosystem. The Gemini Enterprise Agent Platform and Antigravity 2.0 create software lock-in that is independent of the underlying hardware. The direct TSMC relationship secures supply chain independence. And the $175-185 billion in 2026 capex provides the financial firepower to execute this vision at enormous scale.

For Nvidia, the threat is real but not immediate. The company continues to report record revenue, 75% gross margins, and insatiable demand. The CUDA ecosystem, networking business, and next-generation Vera Rubin platform provide substantial competitive moats. However, the structural shift toward custom silicon is irreversible, and Nvidia's largest customers are now its most credible competitors. The question is not whether Nvidia's market share will decline but how quickly and from what base.

For CoreWeave, the risks are more acute. The company's debt-laden financial model, customer concentration, and dependence on Nvidia's ecosystem create multiple points of vulnerability. The current demand environment masks structural fragility. When Google's TPU capacity comes online at scale in 2027, CoreWeave will face a competitor with superior architecture for inference, vastly greater financial resources, and a willingness to invest for market share.

For Nebius, the European sovereign AI narrative provides partial insulation, but the resource asymmetry with Google is overwhelming. The company's partnerships with Nvidia and Meta provide short-term demand anchors, but the long-term competitive dynamics favor vertically integrated hyperscalers that control hardware, software, and distribution.

The AI infrastructure industry is entering a phase of competitive intensity without modern precedent. Five trillion-dollar companies are spending a combined three-quarters of a trillion dollars annually to build the same thing: AI compute capacity. The winners will be determined not by who builds the most capacity but by who achieves the lowest total cost of compute, the most attractive software ecosystem, and the most efficient capital structure. Google has placed credible bets across all three dimensions. Its competitors are now racing to respond.