The chip giant’s massive bet on Baseten signals the AI industry is pivoting from training models to actually running them
Nvidia just invested $150 million in a San Francisco startup most people have never heard of, as part of a $300 million funding round that values six-year-old Baseten at $5 billion. The deal reveals something far more important than another Silicon Valley funding headline. It shows where the chip giant thinks AI’s real money will be made.
While everyone focused on training bigger and bigger AI models, the actual cash is quietly moving to inference: the unglamorous work of running those models billions of times daily to serve real users. Nvidia’s aggressive push into inference startups represents a calculated bet that this operational phase will dwarf the training market it currently dominates.
The funding round, co-led by technology venture capital firm IVP and CapitalG, Google parent Alphabet’s independent growth fund, more than doubled Baseten’s valuation from its September 2025 raise of $150 million. The company has now raised $585 million total, impressive for a firm focused on infrastructure most consumers will never directly touch.
Why Inference Matters More Than Training
Think of training an AI model like sending someone to medical school. Expensive, time-consuming, essentially a one-time capital investment. Inference is that doctor seeing patients every single day for decades. Training happens once or occasionally. Inference is continuous operational spending that accounts for 80 to 90 percent of total AI lifetime costs because every user prompt generates tokens requiring computational processing.
The market numbers tell the story. For the first time ever, inference workloads consumed over 55 percent of AI-optimized infrastructure spending in early 2026, surpassing training costs. Industry analysts project inference workloads will account for roughly two-thirds of all AI compute in 2026, jumping from one-third in 2023 and half in 2025.
The global AI inference market reached approximately $106 billion in 2025 and is racing toward $255 billion by 2030 at a compound annual growth rate near 20 percent. By 2030, inference could represent 70 to 80 percent of AI compute workloads and consume 30 to 40 percent of total data center demand.
This isn’t just reshuffling compute budgets. It’s restructuring the entire AI industry value chain. Companies that built empires selling training infrastructure now face markets where the real volume sits in production deployment. Nvidia sees this inflection point earlier than most and is positioning accordingly.
What Baseten Actually Does
Founded in 2019, Baseten provides infrastructure that companies use to deploy and run machine learning models in production environments. The platform handles the messy operational realities that software engineers face when moving AI from proof-of-concept to serving millions of users: optimized runtime environments, orchestration software, tooling, and infrastructure management that companies would otherwise build from scratch.
CEO Tuhin Srivastava frames the mission simply. “As model-driven products become ubiquitous, we will be the invisible infrastructure behind the AI-first economy,” he said following the funding announcement. The pitch resonates because most companies don’t want to become infrastructure experts. They want to ship products powered by AI.
The customer roster validates this approach. Leading AI companies including Abridge and OpenEvidence run on Baseten’s platform, supporting applications reaching hundreds of millions of users. The company recently partnered with Nebius to provide AI firms with text-to-video inferencing services across the United States, Finland, and France, expanding its geographic footprint and service offerings.
The technical differentiation centers on speed, reliability, and developer experience. Baseten claims the fastest model runtimes, cross-cloud high availability, and seamless workflows powered by what it calls the Baseten Inference Stack. For companies burning tens or hundreds of thousands monthly on inference costs, marginal performance improvements translate directly to budget savings and better user experiences through lower latency.
Nvidia’s Strategy of Investing in Its Own Customers
The Baseten deal continues a pattern raising eyebrows across the industry: Nvidia investing heavily in companies that buy its chips. The company participated in nearly 67 venture capital deals in 2025, surpassing the 54 deals completed in all of 2024, according to PitchBook data. This excludes investments made through its formal corporate venture fund NVentures, which separately engaged in 30 deals versus just one in 2022.
The strategy isn’t subtle. Nvidia backs startups it considers game changers with the explicit goal of expanding the AI ecosystem. By funding companies building on its hardware, Nvidia simultaneously creates demand for its chips while gaining insight into how customers actually deploy AI in production.
Critics point to potential conflicts of interest. Nvidia essentially subsidizing its own customers raises questions about market distortion and whether smaller competitors can access capital on equivalent terms. The company’s dominant market position in AI accelerators combined with its growing venture portfolio concentrates power in ways that could limit competition.
Supporters counter that Nvidia’s investments accelerate AI adoption broadly, benefiting the entire ecosystem. Funding infrastructure companies like Baseten makes deploying AI more accessible to developers who might otherwise lack resources to build comparable tooling. The cycle creates more AI applications, which drives more chip demand, which funds more ecosystem investments.
The regulatory implications remain murky. Antitrust authorities traditionally scrutinize vertical integration where suppliers invest in customers, but AI’s rapid evolution has outpaced regulatory frameworks. As long as Nvidia doesn’t condition chip access on accepting investment or exclusively serving Nvidia-backed companies, the arrangements likely pass legal muster even if they raise competitive concerns.
The New Inference Arms Race
Nvidia’s Baseten investment sits within a broader industry shift where inference optimization has become the new battleground. While training dominated headlines for years, bigger models, more parameters, longer training runs, the operational reality is that trained models must serve exponentially more inference requests to justify their existence.
Deloitte estimates inference workloads will overtake training revenue by 2026, driven by enterprises moving from experimentation to deployment and increasing hybrid and edge deployments. The spending shift lags slightly behind compute allocation because training requires massive upfront capital while inference scales gradually, but 2026 represents the inflection year where inference spending surpasses training across the industry.
Major infrastructure providers are responding aggressively. Lenovo launched three new inference servers at CES 2026, explicitly targeting the transition toward inference-dominant workloads. HPE released purpose-built AI servers optimizing for inference and fine-tuning. Every major cloud provider now offers specialized inference endpoints with pricing models reflecting the different economics of production deployment versus development.
The hardware itself is diverging. Training demands cutting-edge GPUs with massive memory bandwidth, specialized interconnects, and tolerance for higher latency. Inference favors efficiency, lower cost per operation, and minimal latency. This creates opportunities for alternative chip architectures, ASICs, FPGAs, custom accelerators, that can deliver better performance per dollar or performance per watt for inference workloads.
Nvidia’s recent $20 billion acquisition of AI inference startup Groq signals how seriously the company takes this market evolution. Groq developed specialized chips optimizing specifically for inference rather than general-purpose GPU architectures. By acquiring both the technology and executive talent, Nvidia positions itself to serve both training and inference markets with purpose-built hardware.
The Lock-In Problem
Baseten’s integration with major cloud providers, AWS, Google Cloud, and presumably others following this funding, creates powerful ecosystem effects. Once companies deploy their inference infrastructure on Baseten running on Nvidia hardware across these clouds, switching costs escalate dramatically.
The technical integration isn’t trivial. Models optimized for specific hardware and runtime environments require significant engineering effort to migrate. Production systems handling millions of requests daily can’t afford downtime or performance degradation during transitions. Even if alternative infrastructure offers better economics, the risk and effort of migration often outweigh potential savings.
This stickiness benefits Nvidia by ensuring long-term demand for its chips beyond initial training purchases. Companies that train models on Nvidia GPUs and deploy them through Nvidia-backed inference infrastructure become deeply embedded in the ecosystem. Competitive alternatives must offer compelling advantages to justify the switching friction.
The counter-argument suggests this lock-in is unavoidable regardless of Nvidia’s investment. Technical integration challenges exist whether vendors have financial relationships or not. Standardization efforts like ONNX attempt to create portability across platforms, but production deployments inevitably develop platform-specific optimizations that complicate migration.
Can Baseten Justify a $5 Billion Valuation?
The new valuation implies aggressive revenue growth expectations. At typical SaaS multiples of 10 to 15 times annual recurring revenue, Baseten would need to reach $300 to 500 million in revenue to justify current pricing. The company hasn’t disclosed financial metrics, but previous funding rounds and valuation progression suggest significant traction.
The total addressable market supports ambitious targets. If inference spending reaches over $20 billion in 2026 and $250 billion by 2030, capturing even single-digit market share represents substantial revenue opportunity. Baseten competes not just against other inference platforms but against companies building infrastructure in-house, a huge potential market of do-it-yourself engineering effort that could convert to managed services.
The competitive landscape includes both specialized inference providers and generalist cloud platforms offering inference capabilities. Companies like Replicate, Modal, and numerous smaller players target similar customers with varying technical approaches. Meanwhile, AWS, Google Cloud, Azure, and others bundle inference capabilities into broader cloud offerings, competing through convenience and integration with existing services.
Baseten’s differentiation hinges on superior performance, developer experience, and specialized focus. Companies choosing dedicated inference platforms over general cloud services typically prioritize optimization and cost efficiency over convenience. This creates a natural customer segment willing to pay for specialized tooling if it delivers measurable performance or economic advantages.
What This Means for AI’s Future
Nvidia’s investment pattern, Baseten, Groq, OpenAI, Anthropic, Crusoe, and dozens more, sketches a future where the chip giant isn’t merely a supplier but an orchestrator of the AI ecosystem. By funding across the stack from foundational models to deployment infrastructure to application companies, Nvidia positions itself to benefit regardless of which specific technologies or approaches prevail.
This strategic diversification makes business sense. Training may currently drive chip revenue, but inference represents the longer-term sustainable market. Models get trained periodically. Inference runs continuously for years. Companies might negotiate hard on training infrastructure costs knowing it’s one-time expense but accept higher inference costs as operational necessities.
The flip side is dependency risk. If Nvidia becomes essential infrastructure underlying multiple layers of the AI stack, the industry faces concentration that could limit innovation or create vulnerability. Alternative chip vendors, AMD, Intel, numerous startups, struggle to compete not just on hardware performance but against an entire ecosystem of Nvidia-funded companies optimized for Nvidia silicon.
The open question is whether this concentration is temporary or permanent. Historical technology waves suggest market leaders during infrastructure buildouts don’t always maintain dominance once markets mature. Microsoft dominated PC operating systems but missed mobile. Cisco ruled networking hardware but didn’t capture cloud networking. Oracle owned databases but lost ground to cloud-native alternatives.
AI could follow similar patterns. Nvidia’s current dominance reflects the specific requirements of transformer-based models trained with current techniques on today’s datasets. Future AI architectures might favor different hardware characteristics. Inference optimization could fragment across specialized accelerators rather than consolidating around general-purpose GPUs.
Or Nvidia’s ecosystem investments could prove strategically brilliant, cementing platform leadership through network effects that compound over time. By funding the companies building on its platform, Nvidia ensures those companies optimize for Nvidia silicon, creating technical advantages that alternatives struggle to overcome.
The Baseten investment won’t determine these outcomes alone. But it represents a clear signal in Nvidia’s broader strategy of using its current market position and financial resources to shape the AI industry’s evolution. At $150 million, the check is material even for a company Nvidia’s size, especially when considered alongside dozens of similar bets across the ecosystem.
For startups and investors, the message is clear: inference infrastructure is attracting serious capital because the market opportunity is enormous and the shift from training to inference is happening now. The companies that solve operational AI deployment at scale won’t stay invisible for long.
