AI chip supply shortage: plan for 2 years

What just happened—and why I’m paying attention

The AI chip supply shortage is no longer a rumour; it’s been signposted. The Financial Times reports that ASML expects AI demand to exceed supply for two more years (see: Financial Times). Bloomberg.com reports that the chip‑gear maker can’t keep up with AI demand and that the company’s shares rose (see: Bloomberg.com).

I read this as a clear signal: capacity tightness is not a blip. If you rely on accelerated compute to build or run AI, your plan just changed—on value, risk, cost and governance—because scarcity moves from a procurement irritation to a strategic constraint.

What the AI chip supply shortage means for your roadmap

A two‑year imbalance compresses options. Budgets that assumed elastic capacity must accept queueing, higher prices or both. Product roadmaps tied to model training or heavy inference will face hard trade‑offs: delay features, degrade fidelity, or re‑engineer for efficiency.

Risk multiplies. A single dependency—whether a provider, a region or a single large model—can stall delivery. Governance must therefore expand beyond model and data risk into explicit compute risk: who gets capacity, on what terms, and how that aligns to revenue and regulatory commitments.

Value shifts to teams that ship outcomes with fewer GPU‑hours. Features that can be reframed to use smaller models, caching, retrieval or batching will win. Those that can’t may need reprioritisation or different commercial treatment.

Capacity planning in scarcity: lock, diversify, and instrument

Lock what you can on terms you can live with. Multi‑year commitments for training and inference capacity can make sense when scarcity is declared. Structure them with explicit exit ramps, volume tiers and change‑of‑law clauses; you’re buying resilience, not just rate cards.

The readiness questions underneath this are covered in 20 questions to ask before you build with AI.

Diversify on three axes: providers, hardware classes and models. Avoid a single choke point. Multi‑provider tenancy reduces operational risk; model diversity (foundation, distilled, task‑specific) reduces cost risk; and tiered hardware use de‑pressurises peak periods. This is compute capacity planning, not just procurement.

Instrument everything. Treat compute like capital. Tag workloads by product, environment, model and customer SLA. Set budgets in GPU‑hours, with alerts and enforced throttles. Publish a weekly capacity ledger to the exec team; you want visible trade‑offs, not surprise outages.

You don’t need the physics: whether it’s chip‑making acronyms you’ll hear—like ‘EUV lithography’—or datacentre buzzwords, your job is to convert scarcity into prioritised delivery.

Cut GPU‑hours per feature shipped

Start with product design: reduce the default use of the largest models. Use smaller, distilled or domain‑specific models for the bulk of interactions; reserve heavyweight models for true edge cases.

A readiness review before you build surfaces these gaps while they are still cheap to fix.

Optimise prompts and context windows. Truncate aggressively, cache embeddings and responses where acceptable, and use retrieval to keep prompts lean. Every token you don’t send is cost you don’t incur.

Batch and schedule. Move non‑urgent jobs to off‑peak windows; batch inference where latency allows. Align SLAs with business value; not every action needs sub‑second responses.

Exploit compression and quantisation where quality permits. Pilot mixed‑precision and sparsity techniques to cut memory and compute load, and use targeted evaluation to ensure you don’t erode outcomes.

Engineer for locality. Keep data close to compute to reduce transfer bottlenecks and wastage. Eliminate ‘chatty’ service patterns between inference tiers.

Governance matters: make ‘GPU‑hours per feature’ a first‑class KPI alongside revenue impact and user satisfaction. Reward teams that ship more with less.

What to do now

Within 30 days: create a cross‑functional capacity council with authority over AI workload prioritisation and spend. Baseline your current and forecast GPU‑hour usage per product. Tighten GPU procurement discipline with approval thresholds and pre‑negotiated playbooks.

Within 60 days: negotiate multi‑year capacity commitments with clear exit and flexibility clauses. Build a diversification plan across providers and models; pilot a smaller‑model path for your top three GPU‑hungry features.

Within 90 days: implement workload tagging, budget guardrails and automated throttles. Publish a quarterly compute risk register to the board. Tie feature funding to expected GPU‑hour burn and establish a rolling backlog of efficiency improvements.

Keep perspective. The signal from the Financial Times and Bloomberg.com is consistent: plan for scarcity. If you act now—locking capacity wisely, diversifying deliberately and engineering ruthlessly for efficiency—you can keep shipping at pace while others queue.