The cloud giants all flagged the same AI compute crunch in one fortnight.

When OpenAI’s Sarah Friar told Bloomberg on May 1 that the company faces a vertical wall of demand, she handed every management consulting firm an opening. Within five business days, six platform-layer executives printed the same signal: demand outruns compute, capacity is rationed, and the complexity is growing. The leading indicator was explicit, namely that the first compute-capacity strategy briefings would land within fourteen to twenty-one days. Those briefings will be correct about the scarcity. The cure they prescribe will be wrong.

Every major firm has a class of positioning vocabulary that works as lock-in. It sounds like a description of a capability. Its real economic job is to justify staying after the build is done. The test is one question: what does the client own when the engagement ends? KPMG’s operating-model debt is the clearest case, because debt implies perpetual servicing, so the diagnostic becomes the recurring deliverable rather than the rewrite that would eliminate it. Deloitte’s behavior-level governance casts the firm as the monitoring layer after deployment, a subscription that runs on top of the agent. Wipro’s Trust Stack reframes 225,000 bench workers as the human-oversight layer for regulated work. BCG’s tokenmaxxing invents a metric the client does not currently track and positions BCG to benchmark it every quarter. Each pattern is different. The economic function is identical: the client cannot operate the layer without the firm’s ongoing presence.

Compute scarcity makes this vocabulary more defensible than it was six months ago. When AI was abundant, we will manage it for you looked like a visible markup. When AI is rationed, the same line reads as risk transfer. Three mechanics are already in motion. First, scarce capacity creates a procurement-advisory lane these firms are built to occupy, because complex vendor relationships and allocation decisions, exactly what rationed compute produces, are the core offering of ISG, Deloitte cloud practices, and KPMG transformation work. Second, the CFO conversation validates the measurement-framework version of the play: PwC’s Atkinson handed the industry the 7.2x performance-gap number, and that number is the doorway to an assessment, gap analysis, and roadmap sequence PwC owns and runs, made unavoidable because every AI workload now carries an opportunity cost, which is a CFO question.

The third mechanic is the most powerful. The capital-spending-as-layoff-offset move gives the firms a permission structure: Mark Zuckerberg explicitly tied 8,000 cuts announced on May 20 to funding compute. The pitch the firms will run is to cut your operations bench to fund your compute commitment and replace it with our AI-enabled managed service. The client trades fixed staff cost for variable consulting spend, the firm grows revenue, and the work is the same, rebadged through the bench. The whole maneuver converts the client’s fixed labor into the firm’s variable revenue.

The hyperscaler is already collapsing the very layer the firms are racing to occupy. Bedrock managed agents launched on May 5 as an AWS exclusive, with OpenAI, Anthropic, and Amazon Nova running as native managed agents in one platform. The routing logic, fallback chains, and cost-aware load balancing are fixed procedures, and they are becoming native platform features rather than judgment calls that require ongoing consultation. So the question to put to any client weighing the lock-in path is this: when AWS, Google, and Anthropic ship more managed-agent capability natively, which the evidence says happens within sixty to ninety days, does your strategy get cheaper or more expensive? If cheaper, you own the specification the platform executes. If more expensive, you are paying a firm to sit between you and the platform, and that seat is being commoditized in real time.

The fourteen-to-twenty-one-day window is the actionable signal. When the firms publish their compute-scarcity briefings, as the indicator predicts, they will frame the management layer as requiring their ongoing judgment. The counter-frame that lands before that publication owns the vocabulary. The one that lands after merely annotates someone else’s thesis. The platform layer is moving in one direction. The specification that runs on it is the asset.