The apparatus of artificial intelligence today resembles a Difference Engine of unprecedented scale—yet one whose gear tolerances are dangerously mismatched. On one side, the capital expenditure driving this machinery has reached a trillion-dollar scale, with hyperscale operators projected to pour over $1 trillion into infrastructure 3,19 and individual deployments costing $2 billion for hardware alone 22 at a rate of $50 billion per gigawatt 2. On the other side, the per-unit cost of computation—the “token” in modern parlance—has collapsed 500-fold since the original ChatGPT 21, with inference now priced as low as $0.10–$1 per million tokens on flash chips 27. This asymmetry is not a sign of efficient machinery but of a system operating under massive cross-subsidies, with many components operating at a loss for years 2 and core computational components requiring replacement every two years 20. The result is an engine that, despite its immense throughput, risks catastrophic seizure at multiple failure points.
The Gearworks of Capital Expenditure: Trillion-Dollar Tolerances
Examine the primary drive gears. Tesla alone plans to triple its capital expenditures in 2026 1, while cloud backlog has doubled to $460 billion 3,8. Even as hardware advances emerge from NVIDIA, Google, and Amazon, the overall cost of AI operations stubbornly refuses to decline 2. A single terrestrial AI deployment facility demands $2 billion in hardware 22, yet such a facility is engineered to run at a deficit for its first five to six years 2. The operational lifespan of AI compute chips is a mere two years 20, forcing continuous reinvestment into a mechanism that, by design, never depreciates its capital base fast enough. Not surprisingly, Nvidia’s VP of Applied Deep Learning observes that compute costs now significantly exceed employee costs 9, and compute itself has become a scarce resource 4. This is a mechanical monstrosity: a prime mover built from components that wear out before they break even, demanding ever more fuel to maintain output.
The Token Transmission: Asymmetric Friction in Pricing Mechanics
If the capital side is a treadmill, the token economics add a deceptive slippage. Per-unit prices have plummeted: DeepSeek V4 offers equivalent quality at 20% of ChatGPT’s cost 5, Chinese open-source models have been priced 99% below incumbents 6, and inference for DeepSeek’s model costs $0.87 per million tokens against $30 for GPT-5.5 27. Yet every new generation of frontier models is approximately twice as expensive per token to develop 17, and newer tokenizers can consume up to 35% more tokens for identical text 9—a gear ratio that silently increases input costs. Furthermore, the advertised prices are sustained only through heavy subsidies. ChatGPT Plus, priced at $20 per month, actually costs OpenAI an estimated $700 per month per user to operate 10. Subscription plans at $200 per month from Anthropic and OpenAI reportedly provide $8,000–$14,000 worth of compute tokens 12. This is not a market price; it is an engineered loss leader that cannot persist indefinitely, and the machinery will seize unless pricing mechanisms are recalibrated.
Calibration Protocols: Pricing Model Evolution and Seizure Points
The traditional flat-rate subscription—the beloved gear of SaaS—is proving the most brittle component. Heavy users can consume an entire tier’s worth of compute, creating tail risk 23, and loop-protection failures have resulted in single-incident bills of $47,000 16. Startups with limited capital are already abandoning flat-rate models for credit-based systems to avoid insolvency from power users 24. Meanwhile, enterprise pricing for AI coding tools reaches $10,000 per developer per month 7, while open-source alternatives generate entire applications for $2 11. Pegasystems claims its flat-fee model can cut AI costs by over 20× 18, and local large language model inference reduces costs by five orders of magnitude compared to cloud APIs 13. These shifts signal an inevitable architecture change: from predetermined gear ratios to variable transmissions that match cost with consumption, lest the entire revenue mechanism grind to a halt.
Oscillations in the Enterprise Apparatus: Adoption and Hidden Friction
Even in thriving deployments, hidden frictions accumulate. AI workloads incur waste from retries, failed runs, and over-fetching—a 4–5× factor in consumer applications 25,26. A single feature like document summarization can consume 60% of an app’s monthly LLM API spend 28, while for one AI calorie app, Google Maps API costs accounted for 92% of operating expenses 26. Shadow AI usage further destabilizes the system: 56% of employees use unauthorized tools 14, and 44% of organizations have violated their own AI policies 15. This is the noise in the machine—unaccounted, unlicensed, and untamed—threatening the architectural integrity of any enterprise deployment.
Implications for the Salesforce Machinery
For a platform like Salesforce, these dynamics are not abstract; they are the stresses acting on its own subscription-based architecture. The commoditization of base models demands insulation through an orchestration layer that routes queries to the most cost-efficient engine—proprietary, open-source, or local—much like a governor regulates steam pressure to prevent seizure. Salesforce must transition its pricing from rigid seat-based subscriptions to hybrid models that meter consumption without introducing unsustainable friction, capturing value from power users while protecting margins. Moreover, the governance chaos presents an opportunity: by embedding rigorous cost controls, policy enforcement, and model routing as native features, Salesforce can transform these industry-wide vibrations into a stable, differentiated platform. Without such recalibration, however, the machinery risks being undercut by lower-cost assemblies that natively embed cheaper AI, leaving an expensive engine idling as the market races ahead.