Software Revenue Lags Despite Tech Giants’ $292 Billion AI Spend
Silicon Valley is betting the farm on AI. Data centers are straining power grids. Model training costs are heading toward billions. Yet across the software industry, AI revenue remains theoretical.
Hyperscalers – combined with Meta and Oracle — plan to spend $292 billion on AI infrastructure by 2025 – an 88% increase since 2023. Two-thirds of software companies, however, still report decelerating growth in 2024.
Semiconductor stocks have surged 43% year-to-date on AI expectations, while the software index IGV is up 30%. Microsoft, despite its OpenAI investment, has underperformed the IGV by 19% since ChatGPT’s release.
Microsoft’s AI revenue run rate is 3% of total revenue, according to estimates by investment bank Jefferies. Snowflake expects immaterial AI contribution in fiscal 2025. Salesforce isn’t factoring in material contribution from new AI products into FY25 guidance. Adobe’s Firefly AI, launched in March 2023, hasn’t accelerated revenue.
61% of companies report their current architecture cannot support AI workloads without modifications, Jefferies says. Half of data engineers spend most of their time resolving data source connections. Real-time processing and data pipelines remain the primary architectural constraints.
A survey by Jefferies found that large enterprises lead AI adoption due to existing data infrastructure. According to an AWS partner consultant quoted in the report, “a very low percentage of POCs are making it into production.”
Training costs for frontier AI models have reached $100 million. Anthropic’s CEO Dario Amodei projects costs of “ten or a hundred billion” by 2025-2027. Meta states Llama 4’s training will require “almost 10 times more” compute than Llama 3.
Scale AI’s CEO Alexandr Wang reports they’ve “hit a wall on pre-training.” Berkeley’s AI Research Lab data shows scaling large language models yields diminishing performance improvements.
Physical infrastructure limitations also persist. Meta reports power consumption during model training “stretches the limits of the power grid.” Microsoft’s Corporate VP of Azure Hardware states “AI infrastructure can hardly meet the needs of AI model development.”
Microsoft trades at 35.4x FY25 earnings, Meta at 24.6x. The Jefferies report notes companies are shifting from in-house development to purchased solutions. A Salesforce partner reports a customer previously employing “50 people trying to build agents” deployed Salesforce’s AI solution “in a couple of hours.”
Infrastructure spending continues to accelerate while software revenue growth decelerates. The IGV software index trails semiconductor stocks by 13% points in 2024.