AI Compute Wars Enter the ROI Verification Phase

I. The $650 Billion Headline Is a Promise, Not Execution—Earnings Pressure Is More Severe Than It Looks

After the four hyperscalers reported earnings over the past two weeks, markets have largely treated the consensus that "2026 combined CapEx will exceed $650 billion" as proof that the AI buildout is irreversible. But this number has a critical accounting property: it is guidance, not realized spending—and the larger the commitment, the heavier the downstream earnings pressure.

II. Positions That Win Regardless of Who Wins: Four Structural Bottlenecks

If ROI is the real battlefield, the question worth answering is not "which company wins the compute war" but "which suppliers everyone has to buy from." From the verifiable supply chain structure, four positions hold genuine silicon-agnostic neutrality.

Position 1: HBM (High Bandwidth Memory)

SK Hynix holds approximately 62% of the HBM market, with its 2026 capacity fully booked at the start of the year. Micron supplies US-based HBM with 21% market share. Combined capacity from SK Hynix, Micron, and Samsung is fully pre-allocated, with HBM gross margins of 60–70%—far above standard DRAM. NVIDIA Rubin uses HBM4. Google TPU, AWS Trainium, and Microsoft Maia all use HBM. Whoever wins, HBM gets bought.

Position 2: Advanced Packaging (CoWoS)

TSMC's CoWoS packaging makes modern AI chips physically possible. TSMC doubled CoWoS monthly capacity from 70,000 wafers and still cannot keep up; another 30% expansion is planned for 2026. All major custom ASIC programs run on TSMC's 3nm process—a structural reality detailed in our TSM Q1 2026 Deep Research: NVIDIA, Google TPU, AWS Trainium, Meta MTIA, and Microsoft Maia are all TSMC customers.

Position 3: Liquid Cooling and Power Infrastructure

Vertiv's order backlog reached $15 billion as of April 2026—up 109% from two years ago, with a book-to-bill ratio of 1.4x. Eaton completed its $9.5 billion acquisition of Boyd Thermal, marking the end of air cooling dominance and the establishment of liquid cooling as the new standard. Rack power density has climbed from 10kW to over 100kW. Quanta Services (high-voltage transmission) is working through $39.2 billion in backlog. Comfort Systems' data center mechanical and electrical work now exceeds 30% of total revenue.

Position 4: Raw Material Bottlenecks

Helium (used for wafer cooling and leak detection) saw spot prices double after the 2026 strikes on Qatari production facilities (which account for one-third of global supply). Foundries in Taiwan and Korea have begun rationing.

III. The Agentic AI Era: Using Q1 2026 Earnings to Verify Who's Actually Monetizing

"AI agents will drive SaaS growth" is a beautiful narrative, but the gap between story and numbers must be tested through actual financial disclosures. Q1 2026 results have been rolling out and we can begin checking the math.

3.1 Salesforce FY26 Q1: Largest-Scale Validation of AI Commercialization

Salesforce FY26 Q1 (quarter ended April 30, 2025): $9.8 billion revenue, +8% YoY; Data Cloud and AI annual recurring revenue surpassed $1 billion, growing more than 120% YoY; nearly 60% of Q1 top-100 deals included Data Cloud + AI; Salesforce has closed over 8,000 deals since launching Agentforce, of which half are paid. Agentforce has handled over 750,000 requests on help.salesforce.com, cutting case volume by 7% YoY. Data Cloud ingested 22 trillion records in Q1, up 175% YoY.

Reading: Salesforce shows the largest absolute AI monetization scale among SaaS, but the 8% overall revenue growth rate reminds us—growing AI ARR doesn't automatically mean reaccelerating top-line growth.

3.2 ServiceNow Q1 2026: A Beat-and-Raise Quarter, Stock Drops 18% the Next Day

ServiceNow is the most instructive case of this earnings season. Q1 2026 total revenue $3.77 billion (+22% YoY); cRPO $12.64 billion (+22.5% YoY); non-GAAP operating margin 32%; Now Assist customers spending $1M+ in ACV grew over 130% YoY. CEO McDermott raised the 2026 AI ACV target from $1 billion to at least $1.5 billion.

Reading: ServiceNow delivers the most important methodology lesson of this season—growing AI ARR numbers do not equate to share-price appreciation. What the market ultimately prices is the net of "new revenue from AI" minus "AI's impact on the existing subscription model." This logic was already analyzed in our earlier Cloudflare (NET) vs Zscaler (ZS) Battle Research—usage-based pricing is structurally less vulnerable to in-house AI cannibalization than per-seat models, which is the key watershed for SaaS survival in the AI era.

3.3 Snowflake Q4 FY26: Consumption Pricing + AI Monetization Dual Engine

Snowflake Q4 FY26 product revenue $1.23 billion, +30% YoY; RPO $9.77 billion, +42% YoY; 740 net new customers added (a record); 733 customers spending $1M+ on a trailing 12-month basis. Over 9,100 accounts now use Snowflake AI features; Snowflake Intelligence reached approximately 2,500 accounts within three months of launch. Q4 free cash flow margin expanded from 42% to 60%.

Reading: Snowflake's consumption-based model is structurally more aligned with AI workloads than per-seat models. The story is "more AI use → more data queried → more revenue for me"—logically harder to be cannibalized by in-house AI than ServiceNow's seat-based model.

3.4 Palantir: Government + Commercial AI Platform Double-Digit Acceleration

Palantir Q4 2025 revenue $1.41 billion (+70% YoY); US commercial revenue +137% YoY to $507 million; Q4 closed 180 deals of $1M+; US commercial Remaining Deal Value +145% YoY to $4.38 billion; adjusted operating margin 57%, adjusted FCF margin 56%. Rule of 40 score reached 127%. Full-year 2026 revenue growth guidance: 61%.

3.5 Synthesizing 79 Earnings Calls

A systematic review of 79 enterprise software earnings calls yields three conclusions. First, AI is not yet profitable—the goal is "margin neutral." Microsoft, Salesforce, ServiceNow, and nearly every other company described margin pressure from deploying AI products. Second, pricing models are shifting from per seat to per outcome or per agent. Third, copilots are out, and AI that takes action is in—across companies, "AI as assistant" has been de-emphasized in favor of "AI that actually does the work."

IV. Sword and Shield Both Sharpen—Stronger AI Drives Exponential Cybersecurity Demand

This chapter is the inverse of the last. When AI agents can autonomously execute tasks, they can equally autonomously execute malicious tasks. This symmetry transforms cybersecurity from a cost center into a structural growth business—and it is the most underpriced winner of the current AI cycle.

4.1 An Easily Misread Event: Finding Vulnerabilities ≠ Solving Cybersecurity

In April 2026, Anthropic disclosed the capabilities of its frontier model Claude Mythos Preview in controlled testing. Mythos Preview discovered thousands of high-severity zero-day vulnerabilities across every major operating system and web browser, including a now-patched 27-year-old bug in OpenBSD and a 16-year-old flaw. In another test, researchers provided Mythos Preview with 100 Linux kernel CVEs filed in 2024–2025 and asked the model to filter potentially exploitable ones; the model selected 40, then was asked to write privilege-escalation exploits—more than half of these attempts succeeded, and every exploit was written autonomously without human intervention after the initial prompt.

The UK AI Security Institute reached similar conclusions in independent evaluation: under controlled conditions, Mythos Preview executed multi-stage attacks against vulnerable networks and discovered and exploited vulnerabilities autonomously—tasks that typically take human professionals days to complete. On expert-level CTF tasks, Mythos Preview succeeds 73% of the time; two years ago, the strongest model could barely complete beginner-level CTFs.

Bain & Company put it more directly: in 2025, the FBI's IC3 received over 1 million cybercrime complaints with reported losses of $21 billion, up 26% YoY. IBM data shows global average data breach cost has reached $4.4 million, with US-domestic breaches at $10.22 million—an all-time high. Yet most organizations currently plan annual cybersecurity budget increases of only about 10%.

Offense scales exponentially. Defense scales linearly. That gap is the structural demand opening for cybersecurity services.

4.2 AI Agents Are Creating Entirely New Attack Surfaces

If AI just wrote exploits, the cybersecurity market would simply expand quantitatively. But the rise of AI agents introduces entirely new categories of attack surface:

Category 1: Explosive growth of non-human identity. CrowdStrike CEO George Kurtz, on the Q4 FY26 earnings call, noted that 80% of breaches are non-malware-based and tied to identity compromise. CrowdStrike's SGNL AI provides "zero standing privileges" protection covering both human and non-human identities. Every AI agent must now be governed as an identity.

Category 2: Shadow AI. Zscaler's ThreatLabz 2026 AI Threat Report disclosed: enterprise AI usage grew 91% YoY, with customer AI activity sprawling across more than 3,400 applications—quadrupling in the last 12 months alone; data transfers to AI/ML applications increased 93%. When employees paste customer data into AI tools, organizations have no idea where the data flows—precisely the structural reason Zero Trust architecture is being repriced in the AI era, as detailed in our Zscaler (ZS) Research.

Category 3: Prompt Injection and Agent SaaS Attack Surface. After CrowdStrike's acquisition of Adaptive Shield, Falcon Shield's net new ARR grew over 300% YoY—5x since the acquisition closed—because customers are protecting the rapidly growing agentic SaaS attack surface. A single AI agent can span Salesforce, Slack, ServiceNow, and SAP; an attacker only needs to plant a malicious prompt in any one system to weaponize the agent.

4.3 Cybersecurity Companies Use Q1 Earnings to Prove AI Is a Revenue Accelerator

Mapping the structural demand to financial reality:

CrowdStrike FY26 Q4: Ending ARR reached $5.25 billion—the fastest and only pure-play cybersecurity software company to hit this milestone; full-year net new ARR of $1.01 billion (first year exceeding $1 billion); Q4 net new ARR set a record at $331 million (+47% YoY). CEO Kurtz: "CrowdStrike is mission-critical infrastructure—securing AI across every layer from GPU to agent to prompt."

Charlotte AI AgentWorks Ecosystem: At RSA 2026 in March, CrowdStrike launched the Charlotte AI AgentWorks Ecosystem with launch partners including Accenture, AWS, Anthropic, Deloitte, Kroll, NVIDIA, OpenAI, Salesforce, and Telefónica Tech. Note the symmetry: the AI platform winners from the previous chapter—NVIDIA, AWS, Salesforce—are simultaneously cybersecurity ecosystem partners. Cybersecurity is not outside the AI tailwind; it sits at the center of it.

Zscaler FY26 Q1: Revenue +26% YoY to $788.1 million; ARR +26% YoY to $3.204 billion; Rule-of-78 performance. Q1 emerging products (AI security, zero trust, data security) combined ARR exceeded $1 billion. In Q2 FY26, ZDX Advanced Plus (the agentic IT operations product) bookings grew 80% YoY to $100 million.

SentinelOne: Crossed the $1 billion ARR milestone in Q3 FY26 with 23% YoY growth; roughly 50% of quarterly bookings now come from emerging data, AI, and cloud products.

4.4 Why Even Frontier AI Model Companies Are Pushing Cybersecurity Products

The answer to this question is the structural punchline of this chapter—it directly proves the point that "AI finding vulnerabilities ≠ AI solving cybersecurity."

On April 30, 2026, Anthropic launched Claude Security in public beta for Claude Enterprise customers. Built on Claude Opus 4.7, it scans codebases to find vulnerabilities and generates patch instructions. CrowdStrike, Microsoft Security, Palo Alto Networks, SentinelOne, TrendAI, and Wiz are integrating Opus 4.7's capabilities into the security platforms enterprises already run.

In other words, even frontier AI model companies understand: an AI model that finds vulnerabilities must be packaged as a cybersecurity product, deployed, integrated, and operated through existing security infrastructure vendors. The model capability itself is not the security solution—it is a new tool, still requiring platform, workflow, and governance layers.

On the same day, TrendAI (Trend Micro's enterprise AI security business) announced a collaboration with Anthropic. TrendAI's AESIR platform uses Claude Opus 4.7 to "reason like an attacker," autonomously discovering and proving real vulnerabilities; TrendAI Vision One then prioritizes them, maps attack paths, and executes mitigation including virtual patching.

V. NVIDIA Has Prepared a Five-Layer Counterattack

The last commonly overlooked fact: NVIDIA is not passively waiting to be replaced by ASICs. Its counterattack is at least five layers deep.

Layer 1: The CUDA Software Moat

CUDA has over 4 million developers, 3,000+ optimized applications, and is deeply integrated into every major AI framework. It is a self-reinforcing flywheel: developers → framework optimization → hardware sales → R&D reinvestment.

Layer 2: $26 Billion Open-Source AI Investment

NVIDIA plans to invest $26 billion over the next five years developing open-weight AI models—a bet that its hardware lead will hold long enough that even motivated competitors cannot close the gap, while open-model strategy reinforces software dependency.

Layer 3: Direct Integration into AI Agent and Cybersecurity Ecosystems

At GTC 2026, NVIDIA announced collaborations with Adobe, Atlassian, Box, Cadence, Cisco, CrowdStrike, Red Hat, SAP, Salesforce, Siemens, ServiceNow, Synopsys, and other leading SaaS platforms—using NVIDIA Agent Toolkit software to build enterprise AI agents. On the cybersecurity side, NVIDIA is also a launch partner in CrowdStrike's Charlotte AI AgentWorks Ecosystem.

Layer 4: The $20 Billion Groq License—Filling the Inference Gap

Groq is a startup specializing in inference accelerators (LPU, Language Processing Unit). NVIDIA paid $20 billion to license Groq's technology in late 2025 and hired founder Jonathan Ross and President Sunny Madra. At GTC 2026, the first product was unveiled: the Groq 3 LPX rack.

The complete Vera Rubin + LPX system uses NVIDIA Dynamo (the orchestration layer) to manage the division of labor: Dynamo classifies requests, routing prefill and attention to Rubin GPUs, and routing latency-sensitive FFN and MoE decode to LPUs. Paired with Vera Rubin NVL72, the system delivers 35x higher throughput per megawatt for trillion-parameter models compared to the previous Blackwell NVL72 generation.

This is NVIDIA's first dedicated inference hardware. For the past two years, the market has argued "inference will belong to ASICs"; NVIDIA's $20 billion Groq license fills the weakness, converting the topic from "competitor opportunity" into "NVIDIA's own product line." This dynamic was previously analyzed in our Marvell (MRVL) Deep Research: ASIC design services are growing faster, but NVIDIA used Groq to close the inference gap rapidly.

Layer 5: Sovereign AI Opens a New Frontier

Sovereign AI—national-level AI infrastructure—has become a national security issue. Saudi Arabia, Japan, and EU member states are building their own AI capabilities, providing NVIDIA with a customer base diversified beyond US hyperscalers. NVIDIA's business model has shifted from "selling chips" to "delivering complete AI factories": data center (90% of revenue) spans training and inference hardware, networking (Mellanox/InfiniBand), and software (NVIDIA AI Enterprise, NIM microservices, Omniverse).

The metrics researchers should track shift from "how many GPUs sold" to "Groq 3 LPX shipment progress," "Agent Toolkit penetration into SaaS and cybersecurity platforms," "sovereign AI deal visibility," and "market share trajectory of CUDA alternatives."

VI. Methodology: What Should Researchers Do in This Environment?

Connecting the facts from the previous five chapters yields the following work list:

First, reread CapEx beyond the headline. The CapEx figure itself is not the signal; what matters are CapEx-to-FCF ratios, depreciation curves, and the cloud revenue backlog's coverage of CapEx.

Second, shift research focus from chip vendors to neutral bottlenecks. HBM, CoWoS, liquid cooling, power, raw materials—researchers don't need to predict the next winner correctly; they only need to verify that "total demand is still growing."

Third, evaluate AI Agent SaaS stories on the net of ARR vs per-seat impact. The 18% ServiceNow drop reminds us: usage-based pricing (Snowflake, Palantir) is structurally less vulnerable to in-house AI cannibalization than per-seat (Salesforce, ServiceNow).

Fourth, cybersecurity is the most underpriced structural winner of this AI cycle. As AI agents proliferate, the attack surface expands, and cybersecurity budgets must follow. CrowdStrike, Zscaler, Palo Alto Networks, and SentinelOne have already verified this structure through Q1 ARR figures. This is the segment of the AI tailwind that is least priced into current valuations.

Fifth, the NVIDIA narrative requires updating. It is no longer a "GPU sales" story; it is now a seven-pillar platform: Hardware × Networking × Software × Sovereign AI × Agent Toolkit × Groq Inference × Cybersecurity Ecosystem.

Closing Thoughts: Don't Mistake Capital Reallocation for a Trend Reversal

The Q1 2026 earnings season delivered a deeper message than the surface narrative: AI investment has entered the "promises must materialize" phase. $650 billion in CapEx is a promise, $400 billion in depreciation is a known headwind, OpenAI revenue missing targets is a warning sign, hyperscaler free cash flow being squeezed is fact, and ServiceNow's 18% drop is the first SaaS case where the AI monetization story was unbundled.

None of this means AI is over. It means the market is now applying a higher bar to determine who can actually monetize:

• NVIDIA has not lost—five moats are reinforcing simultaneously, and even the inference market (where ASICs were strongest) has been filled by the $20 billion Groq license
• Neutral bottlenecks (HBM, CoWoS, liquid cooling, power) win regardless of who leads
• Strong SaaS players have converted AI into measurable ARR, but watch for per-seat models being cannibalized by their own AI
• Cybersecurity companies sit at the most underpriced position of this AI tailwind—sword and shield both sharpen
• CSPs themselves are entering an ROI scrutiny phase; CapEx upgrades will continue to be punished by markets until cloud revenue growth covers the depreciation curve

The researcher's job is not to predict who wins, but to put these verifiable numbers, visibility, and structural bottlenecks on the table—so that judgment runs slower, steadier, and more grounded in fact than market sentiment. That is what "I teach you how to think, not just what to do" really means.