Frontier AI capabilities and infrastructure demand remain strong, but token price cuts and higher storage costs are new valuation variables for internet stocks

Overview

This report launches BofA's Frontier AI Data Tracker, covering model capability rankings, model usage and spending, token prices, GPU rental prices, and storage prices to assess AI competitive dynamics, valuation sentiment, and margin variables for large-cap internet stocks.

Core views

Anthropic and OpenAI currently define the frontier of large-model capabilities, while Meta's improved model rankings indicate progress at its AI lab. Lower-cost open models have recently gained some usage share, while August API token price cuts have intensified price competition. GPU rental prices remain broadly resilient, reflecting healthy AI infrastructure demand and server useful lives; however, DRAM and NAND inflation continues to pressure cloud and infrastructure margins.

Analysis framework

The report constructs a cross-AI-value-chain monitoring framework using third-party model benchmarks, OpenRouter and Vercel routing data, token cost indices, GPU rental indices, and storage spot prices, interpreted alongside valuations and risk disclosures for large internet companies.

Methodology notes

• Model Capability AssessmentArtificial Analysis Model Index

  Intelligence Index, Agentic Index, and Coding Agent Index

  Uses standardized benchmarks to compare frontier models' overall capabilities, agentic capabilities, and coding-task capabilities.

• Model AdoptionOpenRouter and Vercel Routing Data

  Token Usage Share and Spending Share

  Observes developer adoption, cost-efficiency trade-offs, and commercial traction through requests routed via OpenRouter or Vercel AI Gateway; it does not represent total market usage.

• Inference EconomicsToken Price Index and LLM Token Spend Index

  Blended Cost per Million Tokens

  Tracks standardized cost changes for mainstream models and the actual inference ecosystem to assess price competition and cloud inference economics.

• Infrastructure Supply and DemandSilicon Data GPU Rental Index

  Standardized GPU Hourly Rental Prices

  Covers GPU rental prices across neoclouds, hyperscale clouds, managed platforms, and private rental platforms to observe AI compute demand intensity.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Alphabet

  Benefits from Gemini model iteration, cloud business expansion, and commercialization of AI assets.

  Strengths

  The report believes its strong AI assets, expected double-digit revenue growth, and Cloud margin expansion can support valuation.

  Weaknesses

  The monetization pace and margin impact of integrating AI into search remain uncertain.

  Comparison

  Gemini 3.6 Flash ranks thirteenth in the cited Intelligence Index, behind Claude and GPT-5.6 Sol.

  Risks

  Competitive AI tools diverting search traffic, AI search integration pressuring revenue, DMA compliance pressure, and increased AI capital expenditures reducing free cash flow.

• Amazon.com

  AWS benefits from AI cloud demand but also faces infrastructure costs and potential generative-AI disruption to traffic.

  Strengths

  AWS and advertising provide diversified sources of value, and the report uses a sum-of-the-parts valuation framework.

  Weaknesses

  Rising AI infrastructure and storage costs may compress AWS free cash flow.

  Comparison

  Compared with pure-play model labs, Amazon.com's primary benefit path is cloud infrastructure and enterprise customer demand.

  Risks

  Competition from cloud providers and large retailers, rising AWS costs, agentic AI pressuring direct traffic and high-margin advertising revenue, and macro uncertainty amplifying volatility.

• Meta Platforms

  Improved model capabilities and its open-weight model strategy are important catalysts for its AI competitiveness and valuation.

  Strengths

  Meta MuseSpark 1.2 ranks seventh in the Intelligence Index, indicating progress at its AI lab; the report maintains its Buy view.

  Weaknesses

  High AI investment may weigh on margins in the near term, and advertising revenue concentration remains high.

  Comparison

  Its model ranking still trails leading Claude and GPT-5.6 Sol, but exceeds the Intelligence Index ranking of Gemini 3.6 Flash cited in the report.

  Risks

  Digital advertising's macro sensitivity, AI investment eroding margins, rising fixed assets reducing cost flexibility, AI-native platforms competing for user time and advertising budgets, and regulatory and litigation risks.

Key data

• Frontier Model RankingsClaude Opus 5 ranks first in the Intelligence Index; GPT-5.6 Sol ranks near the topMeta MuseSpark 1.2 ranks seventh in the Intelligence Index, while Gemini 3.6 Flash ranks thirteenth.
• Open Model UsageMiMo-V2.5: 32.8 trillion tokens; DeepSeek V4 Flash: 26.4 trillion tokensOpenRouter month-to-date cumulative data for August.
• Vercel Token Usage ShareDeepSeek 29.7%; Anthropic 24.7%; OpenAI 16.3%; Google 5.1%DeepSeek and OpenAI gained share month over month, while Google and Anthropic declined.
• Vercel Spending ShareAnthropic 64.8%; OpenAI 11.2%; Google 7.8%; Moonshot.ai 6.4%This metric is calculated on a standardized basis using publicly available market prices.
• AI Token Price Index$2.21As of August 13, the report summary states it declined 9% month over month and rose 87% year over year; price cuts and substitution toward more cost-efficient models were the main drivers.
• LLM Token Spend Index$1.15Down 27% month over month in August, versus $1.57 in July.
• GPU Rental PricesB200 $5.63/hour; H100 $2.77/hour; A100 $1.65/hourMonth-over-month changes were -2%, +2%, and flat, respectively; year-over-year changes were +7%, +33%, and +17%, respectively.
• Storage PricesDDR5 approximately +8% month over month; NAND flat month over monthThe two were up approximately 483% and 432% year over year, respectively.

Impact & implications

For large-cap internet stocks, model releases and capability rankings affect AI competitive narratives and valuations. Lower token prices benefit user-side inference costs but may reduce per-unit revenue and margins for cloud services and model providers. Stable GPU rental prices support the AI capital-expenditure demand thesis, while persistently rising storage prices add infrastructure cost pressure.

Risks

• OpenRouter and Vercel data cover only traffic routed through their gateways and exclude direct APIs, cloud platforms, private deployments, self-hosted inference, and internal hyperscale-cloud workloads; they should not be regarded as total market usage.
• Token price competition could outpace demand growth, compressing AI monetization margins for cloud platforms, model providers, and internet companies.
• Continued increases in DRAM and NAND prices could offset some GPU efficiency improvements and raise total AI infrastructure cost of ownership.
• Leading model rankings change rapidly; subsequent releases such as Meta Watermelon or Google Gemini 4 could quickly alter the competitive landscape and market sentiment.
• Large-cap internet stocks remain exposed to company-specific risks, including advertising cycles, regulation, capital expenditures, and traffic diversion by AI products.

What to watch

• Subsequent model releases, capability ranking changes, and open-weight strategies from hyperscale cloud providers and model labs.
• Whether the AI Token Price Index, model price adjustments, and cost-efficiency gains continue, particularly their impact on cloud business margins.
• Changes in usage share and spending share between open and closed models on Vercel and OpenRouter.
• Whether B200, H100, and A100 rental prices remain resilient, validating AI infrastructure demand and server useful lives.
• DRAM and NAND spot price trends and their transmission to cloud infrastructure capital expenditures and gross margins.