The AI boom is no longer one number on one company’s slide. Results released through Tuesday night show demand moving across three linked layers: AMD sold far more data-centre compute, Palantir sold far more data-and-workflow software, and Caterpillar reported the power-equipment demand that increasingly sits underneath the build-out.
That does not make every server dollar, software contract or turbine order “AI revenue.” It does make the chain harder to dismiss as a single-chip story. The useful signal in AI news today is the alignment: compute capacity, operational software and physical power systems are all reporting pressure from the same deployment cycle.
AMD’s data-centre business more than doubled
AMD reported record second-quarter data-centre revenue of $6.7 billion, a 107% increase from a year earlier, driven by demand for EPYC processors and Instinct accelerators. Total revenue reached $11.5 billion, up 50%. The company guided to roughly $13 billion for the current quarter.
The distinction matters: AMD’s data-centre segment includes server CPUs as well as AI accelerators, so the whole $6.7 billion should not be relabelled as generative AI revenue. Yet AI workloads increasingly require both. Training and inference need accelerators; data preparation, retrieval, orchestration and ordinary cloud services still consume CPUs. The latest result suggests the workload is widening rather than remaining isolated inside a small number of frontier-model training runs.
AMD also raised its long-range market expectations, forecasting the data-centre AI accelerator market could reach about $1.4 trillion by 2030. That is a vendor forecast, not an audited future. Buyers and investors should treat it as management’s planning assumption, not as demand already contracted.
Palantir shows the software layer accelerating
Palantir’s overall quarterly revenue rose 93% to about $1.94 billion, according to its results and current reporting. U.S. commercial revenue grew even faster. The company attributes much of that momentum to its Artificial Intelligence Platform, which connects models to governed company data and operational workflows.
This is one of the clearest public signals that enterprise AI spending is moving beyond experimentation. Customers are not merely buying access to a general model; they are paying for the data integration, permissions, deployment and application layer needed to put models inside decisions and processes.
Still, Palantir does not disclose a clean, independently audited split labelled “AIP-only revenue.” Its government and commercial businesses include broader platform contracts. The 93% growth rate is real company revenue growth, but it is not a pure measure of the entire AI software market.
The physical layer is showing up in heavy equipment
Caterpillar reported its first quarter above $20 billion in sales and revenue, while management described strong order rates and a growing backlog. The company sells turbines and power systems used by data centres, and its recent filings have repeatedly linked rising power demand to cloud computing and generative AI.
Caterpillar is not an AI company, and its total quarterly revenue should not be treated as an AI metric. Its relevance is practical: AI automation ultimately runs on electricity, cooling, generators, grid connections and maintenance contracts. When the AI build-out reaches industrial suppliers, deployment schedules become constrained by physical lead times rather than model release calendars.
Model makers can cut token prices quickly. Utilities, turbine manufacturers, construction crews and data-centre operators cannot compress multi-year infrastructure work into a software update. Companies that secure dependable power and integrate compute efficiently may have a more durable advantage than those with the loudest benchmark launch.
What business leaders should read from the numbers
The full-stack demand signal does not justify indiscriminate spending. It raises the standard for capital discipline. Businesses evaluating an AI project should connect four measures: the business process being changed, the software and model cost per completed outcome, the infrastructure capacity required at peak use, and the human review or exception workload that remains.
That is especially important as cheaper inference can increase total consumption. When the price per token falls, teams often put models into more steps, retain longer context and run more evaluations. Unit costs can decline while the total bill rises. Finance teams therefore need cost per resolved case, approved transaction or hours genuinely removed from a workflow—not only API pricing.
The same discipline applies to AI regulation. The EU’s Article 50 transparency rules have applied since August 2, requiring disclosure for many direct human interactions and machine-readable marking for certain generated outputs, subject to defined exceptions and a limited transition for older systems. Capacity and adoption growth do not reduce those duties.
The morning takeaway
The latest AI news is not a new chatbot trick. It is a synchronised demand signal across compute, enterprise software and the equipment that keeps data centres powered. AMD shows the infrastructure layer scaling. Palantir shows customers paying for operational software. Caterpillar shows how quickly the digital boom becomes an industrial order book.
The sober conclusion is neither “AI is all hype” nor “every supplier is an AI winner.” Demand is broadening, but attribution remains messy. The companies that can prove which revenue is AI-driven, which outcomes survive after deployment costs, and which capacity is actually available will define the next phase of AI business trends.