How Silicon Valley's 'tokenmaxxing' is juicing AI demand
Speakers
Deirdre Bosa -- CNBC Host
— CNBC anchor, tech reporter
Summary
- A "tokenmaxxing" trend is emerging where engineers compete to consume AI tokens, creating a volume metric that may not equate to real value, analogous to Amazon's past call center metric failures.
- Eric Glyman (Ramp CEO) notes AI token spend grew 13x over the past year (50% q/q), but CFOs lack budgeting tools, prompting Ramp's launch of AI spend management software to connect token usage to ROI.
- Glyman's data reveals a "K-shaped" outcome: the top quartile of AI spenders (across all industries) more than doubled revenue over 3 years, while the bottom quartile grew only ~12%.
- He argues some "wastage" is a natural part of learning a new technology (like drafts in writing) and that demand may continue growing as cost-per-token falls (Jevons Paradox).
- Glyman observes a shift from frontier models (like GPT-4) to more efficient models (including Chinese models), dropping from >20% to ~4% of token share, signaling a move toward cost efficiency.
- Dan Niles argues the AI investment theme has shifted from a broad, indiscriminate rally to requiring discernment based on cash flow and real ROI differentiation.
- Niles highlights OpenAI's precarious position: it expects to burn $220B through 2029, has ~$20B run-rate vs. $1.4T in capital commitments, and is now cutting prices/projects while Anthropic raises prices/limits access.
- He contrasts OpenAI with Google, which is massively free cash flow positive and can fund its AI ambitions organically, a key reason for their divergent stock performance.
- Niles identifies the rise of "Agentic" AI (e.g., Claude.ai) as a new demand driver, causing token growth to accelerate from ~20% to ~130% in recent months, but expects growth to level out by early next year.
- He posits that the agent-driven AI phase requires orchestration of diverse tasks, which benefits microprocessors (CPUs) over GPUs and increases the need for memory.
- Niles is wary of companies with heavy OpenAI exposure (Microsoft, Oracle) and more positive on those with Google/Anthropic exposure or unique assets (Amazon's logistics, Apple's installed base).
Trade Ideas
Dan Niles states that over half of Oracle's backlog is related to OpenAI, and OpenAI is expected to burn $220B in cash flow through 2029. Oracle's near-term growth is heavily dependent on capital commitments from a partner (OpenAI) that faces severe financial sustainability and competitive challenges, creating significant concentration risk. This high degree of reliance on a financially strained partner makes Oracle's forecasted revenue stream highly uncertain and risky, warranting an AVOID stance. Oracle rapidly diversifies its AI cloud customer base away from OpenAI or OpenAI's financial situation improves dramatically.
Dan Niles states Microsoft owns 27% of OpenAI, which is expected to burn $220B in cash flow through 2029, and that Microsoft's stock is down 20% YTD partly due to this exposure. Microsoft's significant financial and strategic tie to OpenAI creates direct exposure to OpenAI's immense cash burn, questionable path to profitability, and competitive squeeze between Google and Anthropic. This exposure represents a material financial risk and overhang, making Microsoft a less attractive investment relative to other cash-flow-positive AI players. The direction is AVOID. OpenAI achieves profitability sooner than expected or Microsoft successfully insulates its broader Azure business from OpenAI's challenges.
Dan Niles explicitly contrasts OpenAI's massive cash burn with Google being "massively free cash flow positive" and able to fund all its ambitions organically, noting Google's stock is up YTD. In a phase where AI investment requires sustainable capital, Google's robust cash flow provides a significant strategic advantage, allowing it to compete aggressively without the same financial peril as rivals. It is also cited as a winner in the enterprise/consumer squeeze on OpenAI. Superior financial strength and competitive positioning in the key battlegrounds (consumer and enterprise) make Google a relative safe haven and a likely long-term winner, justifying a LONG view. Google fails to translate its financial strength into product leadership or faces severe regulatory action.
Dan Niles states Amazon is hosting Anthropic, has massive physical infrastructure that will benefit from robotics/AI in its e-commerce business, and is the biggest public cloud vendor. Amazon's combination of cloud leadership (hosting a winning AI lab), a cash-generating core business, and unique physical assets that can be optimized with AI provides a diversified and defensible position in the AI buildout. These factors suggest Amazon is well-positioned to capture value across multiple layers of AI demand (cloud, enterprise software, operational efficiency), warranting a LONG view. A severe downturn in overall AI spend or a failure to integrate AI effectively into its logistics operations.
Dan Niles states Apple has a huge installed base of 1.5B iPhone users, has been "so bad on AI" that it can let others invest heavily, and may license AI from Google. Apple's massive distribution and ability to be a "fast follower" in AI by licensing proven technology could allow it to capture AI value with less upfront investment, potentially benefiting in the later stages of adoption. This is a developing, longer-term thesis dependent on Apple's strategic execution. It is not a current catalyst but a future opportunity to monitor, hence WATCH. Apple fails to secure a competitive AI offering or its ecosystem becomes fragmented by rival AI integrations.
Dan Niles states that the shift to "Agentic" AI requires orchestration of many different tasks, which changes the hardware demand from repetitive GPU workloads to more versatile CPU workloads, benefiting microprocessors. The fundamental architectural shift in AI usage (from training/inference to agentic action) drives a change in the primary compute hardware, creating a new growth cycle for microprocessor companies (like Intel) that were previously overlooked. This is a sector-level thematic call on microprocessor and supporting technology companies that stand to benefit from the next phase of AI deployment, warranting a LONG view on the Technology Services sector (specifically the microprocessor niche). The agentic AI trend fails to materialize broadly, or GPU architectures adapt to handle agentic workflows efficiently.
