The race to replace Nvidia's dominance in artificial intelligence hardware is accelerating, with major investors and AI companies placing massive bets on specialized inference chips designed to answer questions faster and cheaper than traditional graphics processing units (GPUs). Chase Coleman III, one of the world's most successful tech investors through Tiger Global Management, recently reduced his Nvidia stake while making Cerebras Systems his 12th-largest holding, signaling a significant shift in how the investment community views the future of AI infrastructure.

Why Are Investors Suddenly Betting Against Nvidia?

Nvidia remains the dominant force in AI training, the process of teaching large language models (LLMs) like GPT-4 to understand and generate text. However, the real bottleneck in AI systems is no longer training; it's inference, the moment when a user asks a question and waits for an answer. This shift has created an opening for specialized chip makers to challenge Nvidia's $5.3 trillion market cap.

Cerebras, a California-based chip startup, has emerged as the leading contender. The company embeds static random-access memory (SRAM), a type of ultra-fast memory, directly onto its chips, dramatically increasing inference speeds. Cerebras claims its new CS-4 system delivers up to 30 times faster inference than Nvidia GPUs on frontier models, though these are company claims that should not be treated as independently verified fact. The company recently went public at a $56.4 billion valuation in May 2026, and its stock has since traded as high as $70 billion to $86 billion.

The power efficiency advantage matters more than raw speed. Data center operators face severe constraints on electricity availability, and Cerebras's promise of delivering better performance per watt could reshape how companies plan their AI infrastructure. The company has already secured 600 megawatts of data center capacity under contract and expects to increase manufacturing capacity more than tenfold in 2026.

Which Startups Are Competing for Inference Market Share?

• Cerebras Systems: Offers wafer-scale chips with embedded memory that claim 30 times faster inference than GPUs; recently went public and has partnerships with OpenAI, Amazon Web Services, and Advanced Micro Devices for hybrid inference solutions.
• Fractile: A British startup in advanced talks to raise $600 million at a $6.5 billion valuation, up sixfold from May 2026; has secured a $250 million chip deal with Anthropic, though chips won't ship until 2027.
• Groq: Pivoted from making its own inference chips to operating a cloud service for fast inference; raised $650 million and now operates 13 data centers globally, with plans to scale toward 200 megawatts by end of 2027.
• Etched: Raised $700 million at a $21 billion valuation this week, building full AI inference systems rather than standalone chips; Jane Street became its first paying customer.

Fractile's valuation jump illustrates investor confidence in the inference market. The company raised $220 million in May at roughly a $1 billion valuation, but three months later, investors are pricing it at more than six times that figure, largely based on a single customer commitment from Anthropic. Fractile's founder, Walter Goodwin, a former PhD student at Oxford, bet that inference speed would become the industry's real bottleneck, and investors are now validating that thesis with capital.

Groq took a different path. Rather than competing directly with Nvidia on chip performance, the company licensed its language processing unit (LPU) technology to Nvidia in a $20 billion deal and is now focusing on operating inference cloud services. Groq raised $350 million at a $3.5 billion valuation in a round that included Nvidia as an investor, signaling that even Nvidia recognizes the value of specialized inference infrastructure.

How Do These Chips Solve the Inference Speed Problem?

The fundamental difference lies in architecture. Nvidia GPUs excel at training because they can perform massive parallel calculations on data stored in separate memory chips. However, during inference, large language models must repeatedly read weights (the learned parameters that define how the model works) from memory as they produce each token, or word fragment. This back-and-forth movement between compute and memory creates latency, the delay users experience waiting for responses.

Cerebras and Fractile take opposite approaches to solve this problem. Cerebras creates massive wafer-sized chips with embedded memory, keeping compute and memory physically close to minimize data movement. Fractile uses an in-memory compute design, storing data directly beside the transistors doing calculations. Both claim dramatic speed improvements, though Fractile's more recent investor materials frame the claim conservatively at 25 times faster and one-tenth the cost, compared to earlier claims of 100 times faster.

These architectural differences come with trade-offs. Cerebras's approach requires special cooling and power management, making it a premium solution. Fractile's chips won't be ready for production until 2027, giving Nvidia and other established players time to respond. Yet the market is moving fast; Etched raised $700 million at a $21 billion valuation this week alone.

How to Evaluate Inference Chip Investments and Claims

• Verify customer commitments: Look beyond marketing claims to actual purchase orders; Anthropic's $250 million deal with Fractile and Cerebras's partnerships with OpenAI and Amazon Web Services indicate real demand beyond hype and investor enthusiasm.
• Compare power efficiency metrics: The most important metric for data center operators is tokens per watt, not raw speed, because power availability is now the limiting factor in AI infrastructure expansion and determines whether a chip can actually be deployed at scale.
• Track manufacturing timelines: Cerebras says CS-4 systems will start reaching customers in the third quarter of 2026, while Fractile's chips won't arrive until 2027, giving different windows for real-world performance validation versus company claims.
• Monitor investor signals: When top-tier investors like Chase Coleman reduce Nvidia stakes while adding positions in Cerebras and Intel, they're signaling belief that the inference market will fragment, with multiple winners rather than Nvidia dominance.

Nvidia is not standing still. The company has acquired Groq's technology through its LPU licensing deal and is developing custom central processing units (CPUs) for agentic AI, systems that can take autonomous actions. Nvidia's next-generation GPU racks will arrive with the advantage of scale, developer familiarity, and a supply chain built around customers already buying by the data center. The company's forward price-to-earnings ratio of just 17 times fiscal 2028 estimates suggests investors still see significant upside.

However, the inference market is large enough for multiple winners. The real test will come when data center operators must choose between familiar Nvidia infrastructure and faster, more efficient alternatives. That decision depends on whether the speed and power gains justify changing procurement habits, rewriting code, and disrupting established relationships. For now, investors are betting that they will.