Yet, a closer examination of Musk’s public commentary reveals a far more complex, often contradictory narrative. Across earnings calls, X posts, and live interviews, Musk has oscillated between showering Nvidia CEO Jensen Huang with lavish praise and issuing stark warnings about the impending obsolescence of Nvidia’s hardware dominance. To understand what Elon Musk has actually said about Nvidia requires tracing a multi-year arc defined by acute supply bottlenecks, astronomical capital expenditures, and Musk’s relentless drive toward vertical integration.
The era of acute dependency and "harder than drugs"
To grasp the gravity of Musk’s remarks, one must first look at the hardware reality of the generative AI boom between 2023 and 2024. As the race to train large language models (LLMs) and vision-based autonomous systems accelerated, Nvidia’s H100 and H200 GPUs became the most coveted commodities in Silicon Valley.
During a May 2023 interview at The Wall Street Journal’s CEO Council Summit, Musk captured the industry-wide desperation in a single, viral quip: "GPUs at this point are considerably harder to get than drugs." The remark was not merely hyperbole; it reflected a severe operational bottleneck that directly threatened the timelines of both Tesla’s autonomous driving ambitions and the nascent launch of xAI.
At the time, Musk frequently emphasized that Tesla’s primary constraint in solving autonomy was no longer talent or algorithmic breakthroughs, but sheer compute capacity. On Tesla’s Q2 2023 earnings call, he openly praised Nvidia’s engineering while lamenting the supply crunch:
"Nvidia has done an incredible job. I have a lot of respect for Jensen [Huang] and Nvidia. If they could deliver us enough GPUs, we might not need Dojo. But they can't. They've got so many customers."
This quote encapsulates the core tension in Musk’s commentary. He has consistently acknowledged the technical superiority of Nvidia’s silicon and its proprietary ecosystem. Unlike many tech executives who publicly downplay their reliance on third-party vendors to project self-sufficiency, Musk has been remarkably transparent about writing multi-billion-dollar checks to Santa Clara. In early 2024, he confirmed on X that Tesla would spend between $3 billion and $4 billion on Nvidia hardware within that calendar year alone, out of roughly $10 billion in total AI-related capital expenditures.
"There is no moat": The strategic shift toward Self-Reliance
Despite his public admiration for Jensen Huang, Musk’s strategic philosophy has always been rooted in first-principles manufacturing and vertical integration. Relying on a single external supplier with gross margins exceeding 70% runs counter to the operational DNA of both Tesla and SpaceX. Consequently, as Musk’s GPU clusters grew into the tens and eventually hundreds of thousands of chips, his rhetoric shifted from grateful customer to aggressive challenger.
The turning point in Musk’s public framing of Nvidia’s long-term dominance came during Tesla’s Q4 2023 earnings call and subsequent X discussions in 2024. While acknowledging that Nvidia held a near-monopoly on AI training hardware at the time, Musk began arguing that this advantage was temporary. Addressing the broader market narrative that Nvidia possessed an unassailable competitive barrier, Musk stated bluntly on X: "There is no moat."
In Musk’s view, the AI hardware market was experiencing a classic gold rush dynamic where Nvidia was selling the best shovels, but competitors, and Nvidia’s own largest customers, were actively designing specialized alternatives to bypass the "Nvidia tax." Musk broke down his skepticism of Nvidia’s perpetual dominance into three primary arguments:
The Inference vs. Training Divide: Musk has repeatedly argued that while Nvidia’s general-purpose GPUs excel at training massive, generalized models in data centers, they are over-engineered and power-inefficient for real-world inference, particularly at the edge, such as inside a moving vehicle.
Power and Thermal Constraints: Throughout 2024 and 2025, Musk shifted the public conversation away from chip shortages toward electricity and transformer shortages. He noted that deploying clusters of 100,000+ liquid-cooled Nvidia H100s and Blackwell B200s pushed local power grids to their absolute limits, creating an urgent economic incentive for custom silicon with superior performance-per-watt.
Custom ASIC Efficiency: Musk contended that application-specific integrated circuits (), stripped of the legacy graphics and general-purpose logic baked into Nvidia GPUs, could eventually outperform Nvidia on both cost and energy efficiency for targeted workloads.
Dojo, AI5, and the critique of General-Purpose Silicon
Musk’s critiques of Nvidia are rarely theoretical; they are almost always tied to his own internal hardware initiatives. For years, Tesla’s Dojo supercomputer, built around Tesla’s proprietary D1 chip, was framed by Musk as a direct hedge against Nvidia. While Musk admitted in 2024 that Dojo was a "long shot" with a low probability of success, he maintained that the potential payoff of breaking free from Nvidia’s pricing power made the multi-billion-dollar gamble worthwhile.
Even more pointed have been Musk’s comparisons between Nvidia’s edge hardware and Tesla’s in-house inference chips (Hardware 3, Hardware 4, and the upcoming AI5). When Tesla pivoted away from Nvidia’s Drive PX platform in 2019 to build its own Full Self-Driving computer, Musk famously declared during Tesla Autonomy Day that it was "objectively the best chip in the world" for autonomous driving, a direct shot at Nvidia’s automotive division.
In subsequent years, whenever Nvidia executives promoted their automotive platforms (such as Nvidia Orin or Thor) as the industry standard for self-driving cars, Musk pushed back. He argued that Nvidia’s automotive chips suffered from the compromises inherent in trying to serve dozens of different automakers with varying sensor suites and software architectures. By contrast, Musk emphasized that Tesla’s custom silicon was co-designed alongside its specific neural network architecture, eliminating wasted silicon area and drastically reducing latency.
The xAI paradox: Criticizing the supplier While breaking deployment records
Perhaps the most fascinating aspect of Musk’s commentary on Nvidia is the stark divergence between his long-term predictions of Nvidia’s commoditization and his short-term execution with xAI.
When Musk founded xAI in 2023 to compete directly with OpenAI, Google, and Anthropic, he could not afford to wait years for custom silicon to mature. Instead, he leaned entirely on Nvidia, and broke industry records in the process. In mid-2024, xAI brought the Colossus supercomputer online in Memphis, Tennessee, uniting 100,000 liquid-cooled Nvidia H100 GPUs on a single unified fabric in just 122 days, a feat that industry analysts previously estimated would take at least 18 to 24 months.
The achievement drew rare, reciprocal public praise from Jensen Huang himself, who called Musk "superhuman" in his engineering execution and noted that no other organization could stand up a cluster of that scale so quickly. Musk responded in kind, thanking Huang and Nvidia’s engineering teams for their intense on-site collaboration to debug networking and thermal bottlenecks at unprecedented scale.
Yet, even as xAI expanded Colossus toward 200,000 H100 and H200 GPUs, and placed massive orders for Nvidia’s next-generation Blackwell architecture, Musk continued to signal that this massive dependency was a transitional phase. In interviews throughout 2024 and 2025, he framed xAI’s rapid Nvidia deployments as a necessary sprint to achieve frontier AI parity, while hinting that future iterations of xAI’s compute infrastructure would inevitably diversify away from reliance on a single vendor’s roadmap.
The Dojo gamble: Designing around the Nvidia bottleneck
Recognizing that reliance on a single supplier posed an existential threat to Tesla’s autonomy ambitions, Elon Musk spearheaded the development of Dojo, Tesla’s proprietary supercomputer built around custom D1 chips. Musk’s public commentary surrounding Dojo often served as an indirect critique of Nvidia’s hardware architecture, not because Nvidia’s chips were underperforming, but because they were generalized.
During Tesla’s AI Day presentations and subsequent investor calls, Musk explained that Nvidia’s GPUs are engineered to handle a broad spectrum of computing workloads, from graphics rendering to large language model training. By contrast, Dojo was designed exclusively for processing massive volumes of video data collected from Tesla’s global fleet. Musk famously remarked that while Nvidia produces exceptional general-purpose AI hardware, a bespoke architecture stripped of unnecessary silicon overhead could theoretically achieve superior performance-per-watt and lower total cost of ownership for vision-based neural networks.
But as the generative AI boom accelerated through 2023 and 2024, Musk’s tone regarding Dojo versus Nvidia evolved from confident disruption to pragmatic hedging. In July 2023, Musk announced that Tesla would spend over $1 billion on Dojo by the end of 2024, while simultaneously confirming that Tesla would continue buying as many Nvidia GPUs as Jensen Huang could deliver. By early 2024, Musk clarified on X that Dojo was a "long shot" worth taking, a high-risk, high-reward venture, while emphasizing that Nvidia hardware remained the primary engine powering Tesla’s Full Self-Driving (FSD) training clusters.
The xAI era and the Memphis supercluster
If Musk’s relationship with Nvidia at Tesla was characterized by strategic tension, his launch of xAI in 2023 transformed him into one of Nvidia’s most vocal champions and largest single customers. Competing directly against OpenAI, Google, and Anthropic meant xAI had no time to wait for custom silicon to mature; raw, immediate compute scale was the only currency that mattered.
In mid-2024, Musk unveiled Colossus, xAI’s massive AI training cluster in Memphis, Tennessee, built around 100,000 liquid-cooled Nvidia H100 GPUs, with plans to double its capacity to 200,000 GPUs using H100 and H200 chips. Throughout the deployment, Musk repeatedly praised Nvidia’s engineering execution and supply chain responsiveness. On X, Musk highlighted how xAI and Nvidia engineers worked side-by-side to bring the 100,000-GPU cluster online in just 19 days, a feat that Jensen Huang publicly lauded as "superhuman."
During this period, Musk’s commentary shifted toward the macroeconomic realities of AI infrastructure:
Capital Expenditure Realities: In Tesla’s Q1 and Q2 2024 earnings calls, Musk estimated that Tesla alone would spend between $3 billion and $4 billion on Nvidia hardware within a single calendar year, pushing total AI-related capital expenditures toward $10 billion.
The GPU Allocation Tug-of-War: In June 2024, reports surfaced, later confirmed and contextualized by Musk on X, that Musk had redirected thousands of Nvidia H100 GPUs originally slated for Tesla to X and xAI. Musk defended the move by stating that Tesla had nowhere to turn the chips on yet because its Austin Gigafactory data center expansion was still under construction, whereas xAI had immediate data center capacity ready to utilize the hardware.
Hardware Lifecycles: Musk frequently noted the brutal depreciation curve of AI hardware, pointing out that today’s state-of-the-art H100 clusters rapidly yield ground to next-generation architectures like Nvidia’s Blackwell (B200) series, forcing companies to continuously reinvest billions just to remain competitive.
Jensen Huang and Elon Musk: Mutual respect between titans
Beneath the transactional dynamics of multi-billion-dollar purchase orders lies a notable rapport between Elon Musk and Nvidia CEO Jensen Huang. Unlike Musk’s famously contentious relationships with other tech leaders, such as Sam Altman, Mark Zuckerberg, or Jeff Bezos, his public exchanges with Huang have remained consistently cordial and mutually complimentary.
The history between the two executives stretches back to the earliest days of modern deep learning. In 2016, Huang personally hand-delivered one of the world’s very first Nvidia DGX-1 AI supercomputers to OpenAI, then co-chaired by Musk, complete with a handwritten note signed by Huang and Nvidia’s engineering team. Years later, Musk frequently referenced that moment when discussing how Nvidia foresaw the deep learning revolution long before legacy chipmakers like Intel or AMD.
When Huang called Musk’s team at xAI "singular in their understanding of engineering and construction" following the Memphis cluster rollout, Musk responded by calling Nvidia’s hardware and CUDA software ecosystem the undisputed gold standard of the industry. Even when discussing competitors like AMD, Cerebras, or Groq, Musk maintained that Nvidia’s entrenched software moat, specifically CUDA, makes displacing the company extraordinarily difficult in the near term.
Power, transformers, and the next bottleneck
As Nvidia successfully ramped up production of its Hopper and Blackwell architectures throughout 2024 and 2025, Musk’s warnings about AI infrastructure bottlenecks shifted away from silicon and toward the physical power grid.
Speaking at the Bosch Connected World conference in early 2024 and reiterating the point across numerous interviews into 2025 and 2026, Musk observed that the primary constraint on AI scaling had transitioned from "chip-constrained" to "voltage-transformer-constrained" and "electricity-constrained."
In Musk’s assessment, Nvidia had solved the compute density problem so effectively that data centers can now pack unprecedented processing power into a single server rack, creating an insatiable demand for gigawatt-scale power generation and step-down transformers. Musk famously quipped that while everyone was scrambling to buy Nvidia GPUs, very few executives were paying attention to the mundane electrical equipment required to actually plug them in.
The verdict on Nvidia’s dominance
Across more than a decade of public statements, earnings calls, and social media posts, Elon Musk’s perspective on Nvidia has evolved from casual recognition of a useful graphics vendor to awe at an indispensable infrastructure monopoly.
While Musk continues to invest in proprietary silicon through Tesla’s Dojo and AI5/AI6 inference chips, driven by his core philosophy of vertical integration and cost reduction, he has consistently acknowledged that Nvidia remains the foundational bedrock of frontier AI training. For Musk, Nvidia is both an essential partner enabling his most ambitious ventures and the ultimate benchmark that every custom silicon project must beat to justify its existence.
