Nvidia CEO Jensen Huang defended heavy AI infrastructure capex at Goldman Sachs conference, arguing that 100x returns justify continued aggressive investment.
On the third day of Goldman Sachs’ annual Communacopia + Technology Conference, Nvidia CEO Jensen Huang addressed market concerns regarding the company’s $500 billion financing platform, which critics have labeled “circular financing.” He countered with a direct financial comparison: “We put in $1 and get $100 back—is that circular financing? If that counts, then we’ll do more of it.”
The exchange quickly became one of the conference’s most discussed moments. In a subsequent interview with Goldman Sachs semiconductor analyst James Schneider, Huang outlined the drivers behind AI infrastructure’s rapid expansion. He reiterated projections that the AI infrastructure market will reach $3 trillion to $4 trillion by 2030 and clarified that Nvidia’s targeted 70% year-over-year revenue growth for fiscal 2027 is limited by supply constraints, not demand.
This theme was echoed across the conference. Executives from SpaceX, CrowdStrike, Palo Alto Networks, and other major technology firms consistently emphasized that compute power, electricity, and data center capacity—not end-user demand—remain the primary bottlenecks constraining AI growth. These constraints are projected to persist through 2028.
A Direct Confrontation on the Circular Financing Controversy
The controversy traces back to a major initiative announced in August, when Nvidia partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR, and other financial institutions to launch a compute financing platform. The initiative aims to mobilize over $500 billion in third-party capital to help AI model developers build infrastructure. Critics argue the model blurs the line between investment and sales: Nvidia funds AI companies or cloud providers, who then use those funds to purchase Nvidia hardware. This dynamic mirrors practices that contributed to the downfall of suppliers like Lucent Technologies during the prior internet infrastructure boom. By linking third-party capital, the financing platform, and hardware procurement, the structure forms a closed loop. Huang’s rebuttal directly addressed the final transaction in that cycle.
Huang confronted the issue head-on, using financial metrics as his primary defense. “This isn’t circular financing, because we put in only a little money and get back a lot of capital,” he stated. He stressed that Nvidia rigorously validates each investment, ensuring partners hold verifiable customer contracts capable of generating independent revenue. “I’ve seen contracts totaling $100 billion,” he noted. “I’m a fully informed investor. I don’t take risks. What I need are near-certain investments.” He characterized the arrangement as a “very smart strategy” designed to expand distribution channels for Nvidia’s technology architecture. Goldman Sachs research supported this view, noting that Nvidia’s financial commitments serve as ecosystem seed capital and that the broad return potential and diversified investor base fall outside the traditional definition of circular financing.
70% Growth Is a Supply Ceiling, Not a Demand Ceiling
During the interview, Huang attributed AI infrastructure’s rapid expansion to two converging industry shifts: the transition from retrieval-based to generative computing, and the effective end of Moore’s Law. He explained that for the past six decades, computing was largely “pre-recorded”—users queried stored data, and algorithms returned preference-based results. Generative AI, by contrast, demands real-time, context-aware synthesis, requiring a fundamentally new, continuously active computing layer. Simultaneously, traditional transistor scaling has plateaued, meaning performance gains no longer stem from smaller process nodes. Nvidia’s countermeasure is aggressive co-design via CoWoS multi-chip packaging and NVLink scale-up architecture. “If you don’t have NVLink, if you can’t achieve truly excellent scale-up, you’re dead, because Moore’s Law is your enemy,” Huang declared, prompting strong reactions from the audience.
He reiterated that Nvidia’s 70% year-over-year revenue growth target for fiscal 2027 represents a supply ceiling, not a demand ceiling; without hardware constraints, growth would be substantially higher. Analysts project Nvidia’s current fiscal year revenue at approximately $400 billion, which would rise to roughly $680 billion under the 70% growth scenario. “We can achieve 70% year-over-year growth, and we’re confident about that,” Huang said, adding that he plans to “eat more Korean fried chicken” and attend more supplier dinners to strengthen supply chain partnerships. Drawing a parallel to TSMC’s role in semiconductors, Huang noted, “I can build my company on top of TSMC; you can build your startup on top of Nvidia.” He emphasized that every AI laboratory can develop its products on Nvidia’s platform, functioning “somewhat like TSMC.” Acknowledging widespread bottlenecks in upstream packaging, wafers, memory, and downstream data center power and real estate, Huang highlighted Nvidia’s ecosystem scale and customer visibility as key navigational advantages. “We are tracking every gigawatt of land, power, and shell on the planet—literally everything on Earth.”
SpaceX: From Rocket Operator to AI Infrastructure Platform
SpaceX CFO Bret Johnsen delivered another major highlight, clarifying that the company is positioning itself as an integrated AI infrastructure platform rather than solely a launch or satellite operator. Its vertically integrated stack spans rockets, power generation, compute infrastructure, AI models, customer networks, and Starlink connectivity. Key metrics disclosed underscored an aggressive growth trajectory: compute revenue sits at the upper end of a $30 billion to $50 billion range per gigawatt; the company targets $100 billion in annualized total revenue by year-end; compute capacity will exceed 2 GW by the end of 2026 and reach 5–10 GW by the end of 2027; AI satellite deployment is scheduled for 2027; and Starlink Mobile will launch in the first half of 2028. Johnsen noted that a newly signed compute licensing agreement earlier this month strengthens confidence in hitting the $100 billion annualized revenue target. Regarding orbital compute, SpaceX plans to deploy AI satellites in 2027 and anticipates that space-based compute costs will reach parity with terrestrial infrastructure faster than previously forecast. By integrating AI models, ground-based compute, Starlink, Starship, and orbital capabilities into a single growth narrative, SpaceX aims to establish a distinct AI infrastructure pathway separate from traditional cloud providers.
Cybersecurity: The Most Underappreciated Segment of the AI Era
Cybersecurity emerged as the most frequently discussed sector at this year’s conference, with Huang, Palo Alto Networks CEO Nikesh Arora, and CrowdStrike CEO George Kurtz each highlighting the segment’s strategic importance from distinct perspectives. Huang outlined a straightforward causal chain: AI accelerates automated coding, which simultaneously speeds up code exploitation and attack vectors, thereby surging demand for vulnerability remediation. “What better way to create demand than to create a problem?” he remarked, noting that cybersecurity demands continuous operation and will unlock substantial new revenue streams for major AI laboratories. Nvidia is currently collaborating with CrowdStrike, Cisco, and Palantir to develop red-team and blue-team defense systems powered by its Nemotron models.
Arora highlighted that AI is rapidly advancing both vulnerability discovery and attack sophistication, making integrated security platforms essential over fragmented vendor stacks. With roughly 180 million deployed sensors processing 19 petabytes of data daily, Palo Alto Networks has established a formidable data and training moat. “The AI security industry hasn’t even been built yet,” Arora stated, adding that $5 trillion in compute investment over the next five years will generate massive network traffic expansion—all of which will necessitate advanced security inspection.
Kurtz emphasized that CrowdStrike’s decade-plus repository of proprietary, labeled security data serves as a core competitive advantage. Its domain-specific AI model reduces remediation costs to approximately $0.03 per activity, compared to roughly $10 for frontier general-purpose models, while maintaining comparable efficacy. Early demand for its new Guardian product has been robust, with Flex agreements reaching 935 deals and $2.3 billion in flexibility commitments; quarterly Flex transaction volume more than doubled last quarter.
Despite their critical roles, Palo Alto Networks and CrowdStrike hold a combined market capitalization of approximately $490 billion—a notable discount relative to trillion-dollar AI peers. This valuation disparity has intensified market focus on the cybersecurity sector’s potential for multiple re-rating.
Semiconductor Equipment and Memory: Supply-Demand Remains Tight
Lam Research (LRCX) CFO Doug Bettinger struck a firm tone at the conference, sa