NVIDIA Is No Longer Just Selling Chips
What NVIDIA's $18.6 billion in Q1 private investments, Marvell's preferred stock deal, and Palo Alto's $21 billion identity acquisition reveal about who controls the next layer of AI infrastructure
NVIDIA deployed $18.6 billion into private companies and infrastructure funds in the first quarter of fiscal 2027. Some of them are AI model makers that, by NVIDIA’s own description, “may indirectly use” NVIDIA products. That phrase is careful for a reason. What it describes is demand engineering: NVIDIA is investing in companies whose success creates demand for NVIDIA chips. At the same time, Marvell Technology, which designs custom AI processors and high-speed connectivity hardware for hyperscalers and cloud providers, disclosed that NVIDIA invested $2 billion in Marvell preferred stock as part of a strategic partnership linking Marvell’s custom processors with NVIDIA’s infrastructure ecosystem. NVIDIA is now an equity holder in one of its own supply chain partners.
The consensus view of NVIDIA is a demand story. Hyperscalers are spending without restraint and NVIDIA is capturing the bulk of it. That framing is accurate as far as it goes. But it describes NVIDIA as a passive beneficiary of decisions made by others. Recent company filings describe something different: a company actively building the financial architecture of its own demand. The $18.6 billion is not passive portfolio management. It is a capital strategy to keep the AI ecosystem dense, funded, and dependent on NVIDIA hardware. That is a different kind of competitive advantage than the one most investors are pricing.
NVIDIA grew 85 percent year over year to $81.6 billion in quarterly revenue. Data center revenue, at $75.2 billion, is now split roughly evenly between hyperscalers and a newer category NVIDIA is calling ACIE, which stands for AI clouds, industrial, and enterprise customers. That split is worth thinking about. The hyperscaler spending story dominated the AI buildout narrative for two years. Enterprise and industrial AI is now equally large in NVIDIA’s own numbers, and the company reorganized its entire reporting structure this quarter to reflect that shift.
NVIDIA’s Leverage Is No Longer Just Technical
The number to watch from NVIDIA’s quarter is not $81.6 billion. It is $15.9 billion, the unrealized gain on NVIDIA’s publicly-traded and non-marketable equity portfolio in a single quarter. That investment book produced gains equivalent to roughly 20 percent of the quarter’s total revenue. Management offered minimal narrative about it in the report. The composition of the portfolio is not disclosed. The duration is not disclosed. What is clear is that it can reverse quickly, and when it does, NVIDIA’s reported net income swings hard.
This investment activity is not incidental to the business. NVIDIA’s Q1 operating income alone was $53.5 billion — the company generates cash faster than almost any enterprise in history, and it is deploying that cash into the ecosystem that buys its chips. The Marvell deal exemplifies this. Marvell designs custom AI chips for companies that want specialized processors rather than NVIDIA’s standard products. That positioning puts Marvell in at least theoretical competition with elements of NVIDIA’s own networking and compute business. By taking a $2 billion preferred equity stake, convertible into approximately 21.8 million Marvell common shares, NVIDIA changed the relationship from arm’s-length to financially intertwined. Marvell’s success now benefits NVIDIA as an equity holder. And Marvell’s products get endorsed by the imprimatur of an NVIDIA strategic partnership.
Think about what Intel Capital was doing in the 1990s. Intel invested in software companies and content producers because it understood that chip demand is downstream of the applications running on chips. More compelling software meant more chips sold. Intel’s investments were not primarily about financial returns. They were about ecosystem construction. NVIDIA’s position in 2026 is structurally similar but scaled up by an order of magnitude. Intel was deploying hundreds of millions to perhaps a billion or two annually at its peak. NVIDIA deployed $18.6 billion in a single quarter without visibly constraining its balance sheet. The strategy is the same. The ambition is different.
Three direct customers now represent 21 percent, 17 percent, and 16 percent of NVIDIA’s revenue, and one unnamed entity, described only as “an AI research and deployment company,” contributes a meaningful amount of revenue through cloud purchases without appearing as a direct customer. That anonymized concentration disclosure is worth paying attention to. Indirect customers purchasing through cloud intermediaries are sometimes described in category terms rather than by name, but the specificity here, an unnamed company called out as contributing a “meaningful amount” of revenue, is a signal worth tracking. If that entity turns out to be a company in which NVIDIA has also invested, the relationship is more complex than a standard vendor arrangement, and the revenue is more concentrated than it appears.
NVIDIA also acknowledged that expanding energy capacity for AI data centers is a “multi-year process” with significant regulatory, technical, and construction challenges. That is the company saying, explicitly, that its forward revenue depends on someone else’s ability to build physical infrastructure at speed. The $18.6 billion in investments is, among other things, a way to ensure that the companies at the leading edge of AI deployment stay well-funded enough to keep building.
Every AI Agent Needs a Permission Slip
Palo Alto Networks spent $21.1 billion to acquire CyberArk in February 2026. CyberArk sells identity security software: workforce identity management, privileged access management for administrative accounts, machine identity management for non-human systems, and identity governance across enterprise environments. The operating impact was immediate and substantial. Palo Alto’s operating margin swung from positive 9.6 percent in the year-ago quarter to negative 6.1 percent in Q3 fiscal 2026, driven by amortization of CyberArk’s acquired intangible assets, accelerated equity vesting, cloud hosting cost increases, and integration expenses. Gross margin compressed 520 basis points in a single quarter.
So why identity, and why now.
The answer is in how an enterprise deploys AI agents at scale. An AI agent that can approve a payment, query a production database, or send communications on behalf of a human needs authentication credentials the same way a human does. It needs to be verified as an authorized entity, with defined permissions, auditable behavior, and access that can be revoked. Traditional identity systems were built for people at keyboards. They assume a human logging in from a known location, following recognizable usage patterns. AI agents do none of that. They operate continuously, at machine speed, often without human supervision, and their behavior patterns are difficult to characterize in advance.
Okta, which manages identity and access controls for corporate users across enterprise environments, is building what it calls non-human identity products, meaning the authentication infrastructure specifically for AI agents. The products are currently in development and early access. At 11 percent subscription revenue growth, the market is not yet paying for the option. But the company’s direction is unambiguous, and so is the logic: every AI agent that touches enterprise data is a new identity management problem.
CrowdStrike, which protects endpoints, cloud workloads, and enterprise identities from a unified sensor platform, extended its coverage in the most recent quarter to include what it calls the prompt and agentic interaction layer, the interface through which AI agents communicate with enterprise systems. CrowdStrike’s ARR, which is the annualized value of active subscription contracts, grew 24 percent to $5.5 billion, and the company reported its first quarterly profit: $27.8 million versus a $104.3 million loss a year earlier. The recovery is genuine. What management was equally direct about is the persistent shadow of the July 2024 software update incident that caused millions of Windows machines to fail globally. The company is still working through the commercial consequences: elongated sales cycles, customer packages that bundle discounting with service credits and extended payment terms, and higher contraction rates as customers stretch subscription terms rather than renew at standard prices.
Zscaler, which inspects network traffic between corporate users and applications to enforce security policy, grew revenue 25 percent in the most recent quarter and increased R&D spending 34 percent over the first nine months of its fiscal year against 26 percent revenue growth over the same period. Operating losses widened. Free cash flow margin held at 29 percent. Management acknowledged directly that customers are requiring more approvals for large purchases and that deal timelines are longer. The investment ahead of revenue reflects conviction that the security market is restructuring around AI. The pace of that restructuring is moving more slowly than the investment thesis requires.
The convergence point across these four companies is the identity layer. For twenty years, enterprise security was organized around the perimeter: inspect traffic at the network edge and trust what comes from inside. AI agents break that model. The relevant security question stops being where the request is coming from and becomes who or what is making the request. Palo Alto’s $21 billion bet is a claim that the company controlling identity for AI agents controls a significant portion of enterprise security spend for the next decade. The question is whether the market agrees before the integration costs force clarity on whether that claim was right.
What to Watch
The test for NVIDIA’s ecosystem investment strategy becomes visible when the identities of portfolio companies come into focus. The anonymized concentration risk disclosure, naming one significant indirect buyer without identifying it, is worth tracking in future filings. If that entity turns out to be a major NVIDIA portfolio company, the revenue concentration and the investment strategy are more intertwined than the current narrative implies. Any voluntary disclosure of indirect customer identity, or any disclosure linking a portfolio investment to a material revenue relationship, changes the picture meaningfully.
For Palo Alto Networks, the clarifying metric is subscription and support gross margin, which compressed 530 basis points to 66.5 percent in Q3. Recovery toward prior-period levels over the next two quarters would indicate the costs are temporary integration drag. If margin stays compressed, it signals that CyberArk’s cloud delivery architecture carries structurally higher costs than the legacy Palo Alto business, which changes the platform margin thesis over a multi-year horizon.
For the enterprise software cohort, Workday’s explicit disclosure about reduced headcount-level commitments at renewal and GitLab’s Dollar-Based Net Retention Rate, a measure of expansion revenue within existing customers, declining from 122 percent to 117 percent set up a two-quarter thesis test. If AI-driven efficiency is genuinely shrinking the number of software seats customers need, both metrics should continue to deteriorate across the platform-subscription category. If the compression is concentrated in government, healthcare, and higher education tied to federal funding uncertainty, as Workday’s own explanation suggests, those metrics should stabilize as that uncertainty resolves. Two more quarters of data distinguish a sector-wide structural shift from a vertical-specific cyclical one.