OpenAI, Oracle & AMD shake up AI

The Shifting AI Infrastructure Landscape: Oracle and AMD Challenge Nvidia’s Dominance

A significant shift is underway in the AI hardware market, with an emerging cluster of competition forming around OpenAI, AMD, and Oracle, directly challenging Nvidia’s long-held dominance. The discussion centers on several key deals: Oracle’s plan to deploy 50,000 AMD chips, OpenAI’s major deal with Oracle, and OpenAI’s investment and partnership with AMD. This creates a “troika” that diversifies the AI chip supply away from a single source.

While this circular flow of investment—where major players invest in each other—raises questions about a potential “AI bubble,” the panelists argue it’s different from the dot-com era. The productivity gains from AI are nearly instantaneous, with a 43% increase since ChatGPT’s introduction, whereas the dot-com bubble’s returns took decades to materialize. The current build-out of data centers is seen as consistent with forecasted demand.

A key part of the discussion focused on Nvidia’s primary competitive advantage, or “moat”: its proprietary CUDA software platform. The consensus is that this moat is becoming “much more shallow than we think.” There are two main reasons for this:

1. Compatibility Layers: Competitors like AMD are building compatibility layers for their hardware (e.g., ROCm), effectively standardizing against CUDA.
2. Abstraction Layers: Companies like OpenAI are aggressively developing programming abstractions like Triton, which hide the underlying GPU architecture and make applications hardware-agnostic.

The panel notes that while hardware from AMD and Nvidia is reaching near-parity, the new bottleneck is shifting from silicon to physical infrastructure, particularly energy and power. The demand for power is growing at a rate that the traditional energy sector cannot match. This will likely lead to very aggressive silicon replacement cycles, as it’s cheaper and faster to swap in a more power-efficient chip than it is to build a new power plant.

Evaluating the Competition: The CAISI Report on DeepSeek

The conversation shifted to a recent report from the Center for AI Standards and Innovation (CAISI), a unit within the U.S. Department of Commerce. The report evaluated the open-weight Chinese model, DeepSeek, and concluded that it lags behind U.S. models in performance, cost-effectiveness, and security.

However, the panel offered a nuanced interpretation:

• Benchmarks vs. Standards: The report highlights the difference between performing well on public benchmarks versus meeting robust standards. DeepSeek appears to have “overfit on the public common benchmarks” to make a splash. In contrast, U.S. companies often use proprietary, black-box evaluations that test for a wider range of capabilities and weaknesses. Aaron Botman compared it to the difference between a car’s top speed (a benchmark) and its crash safety rating (a standard).
• True Cost of Ownership: The report revealed a critical insight into model economics. While DeepSeek may have a lower per-token cost, its per-task completion cost was found to be significantly higher. This means that for a given enterprise task, the model required more resources, making it less cost-effective in practice.
• Mindshare for Open Models: Despite the report’s findings, DeepSeek saw a 1000% increase in downloads. This signals a strong and growing appetite in the market for open-weight models, driving enterprise behavior and increasing the demand for choice beyond proprietary systems.

The Frontier of Open Models: Reflection AI’s $2B Gamble

The panel discussed Reflection AI, a startup founded by former DeepMind alumni that raised $2 billion to become the “leading frontier AI open source company for the United States.” The panelists expressed skepticism, suggesting the market for new foundation model companies is in its “tail end.”

The key challenges for Reflection AI include:

• Saturated Market: “Just being open source is insufficient” as a differentiator in today’s market, with established players like Meta and even IBM offering powerful, transparently developed open models.
• Execution Risk: The company has an ambitious roadmap but has not yet released a single model.
• “Open” is Nuanced: Reflection AI plans to be “open weights” but not “open training,” meaning the data and methods used to build the model will remain proprietary. This limits true transparency and auditability.

The discussion concluded that while the space for generic large language models may be saturated, opportunities remain in verticalization—creating specialized models for specific enterprise domains like finance or healthcare. There is also a potential for a “Foxconn for models” business model, where a company provides the core technology and pre-trained base for other companies to build their own proprietary models at a reasonable cost.

Automating Venture Capital: Hype vs. Reality

The final topic was a story about a small, $75 million venture capital fund, DVC, that claimed to have “fired all its analysts” (five people) and replaced them with AI agents. The panel largely viewed this as a marketing narrative rather than a fundamental shift.

The reality is that this is a story of job transformation, not job replacement. The value of human workers is amplified by AI tools, leading to new roles like “AI translators” and “agent orchestrators.” In the context of a small, early-stage VC fund, success is driven by relationships and deal flow, not deep quantitative analysis. The “AI analyst” is likely a tool to help manage and perform initial diligence on deals sourced through the fund’s network of Limited Partners (LPs). At this stage, investment decisions are more of a “crapshot” based on the quality of the founding team, where human judgment remains paramount.

• Tags:
• Venture Capital
• Benchmarks
• AI Safety
• CUDA
• AI Infrastructure
• Nvidia
• AMD
• Open-Source AI
• DeepSeek
• Oracle
• CAISI
• Reflection AI