Enterprise AI in 2026 is not a contest between frontier models and local ones, or between NVIDIA and commodity hardware. It is a placement problem: five workloads — pretraining, fine-tuning, inference, R&D and agentic systems — spread across a stack that now runs continuously from a 230 kW rack to a desktop, under three constraints that do not negotiate: silicon supply, electrical power, and jurisdiction.

NVIDIA sells a platform, not a chip

Vera Rubin is six co-designed chips: 288 GB of HBM4 per GPU at 22 TB/s, 50 PFLOPS of NVFP4 inference, NVLink 6 at 3.6 TB/s per GPU. An NVL72 rack holds 20.7 TB of HBM4, delivers around 3.6 EFLOPS, and draws an estimated 190–230 kW under full liquid cooling, at a reported $5–7M.

The durable value sits above that silicon: CUDA-X, TensorRT-LLM, Dynamo — which splits compute-bound prefill from bandwidth-bound decode and is often worth more than a hardware generation — NeMo, NIM microservices, Run:ai scheduling, and supported AI Enterprise builds. The same stack runs on a DGX Spark (128 GB unified memory, 273 GB/s, inference to 200B parameters, 240 W) and on Jetson at the edge.

The ecosystem position is wider still: roughly $2B each into CoreWeave and Nebius plus a $6.3B demand backstop, about $30B into OpenAI and $10B into Anthropic, and $12.93B for Hugging Face — the distribution layer for nearly every open-weight model an enterprise would run locally. NVIDIA has committed to keeping the Hub open and multi-accelerator. Sensible programmes will mirror weights internally and pin revisions regardless.

Frontier models earn the hard tail

GPT-6 Astra and Claude Fable 5.1 list at $10/$50 per million input/output tokens; Gemini 3.1 Pro at $2/$12, Grok 4.7 at $2/$6. What that buys is not tokens. It is a capability level that improves every few months with no work from you, plus tool-call reliability across long trajectories, context that degrades gracefully, and safety scaffolding you would otherwise build yourself.

Consuming it need not forfeit sovereignty. There are four levels: in-region endpoints under zero-retention terms; dedicated capacity with private link and customer-managed keys; weights you hold and can reproduce; and a fully domestic stack. Most organisations that say they need the fourth need the second or third, and could have it this quarter.

Most volume is not frontier work

Gemma 4 spans 2.3B effective parameters for phones up to a 26B mixture-of-experts with 3.8B active; Phi-4 at 14B runs in about 10 GB; Llama 3.2 3B fits in 2 GB and still handles tool calls competently. For classification, extraction, routing, summarisation and retrieval-grounded answering, the gap against a frontier model is narrow — and it is measurable on 200 graded examples from your own traffic, which is the only evidence worth acting on.

The non-cost advantages matter more than the savings. A pinned local model behaves identically a year later, which audited decisions require. It answers in tens of milliseconds. It works during an outage. The data never leaves the device. Where open models still lag is long agentic trajectories and out-of-distribution reasoning — design for that gap rather than pretending it has closed.

Apple made 512 GB a commodity purchase

The M5 Ultra Mac Studio offers up to 512 GB of unified memory at 1.2 TB/s from $5,499. Decode is bandwidth-bound — tokens per second is roughly bandwidth divided by active parameters times bytes per parameter — which puts a 70B model at 4-bit near 14 tok/s and a 120B MoE near 43. New Neural Accelerators in each GPU core address prefill, historically Apple's weak point, with a claimed 9.8x faster prompt processing than the M1 Ultra.

Per TB/s of memory bandwidth, a Mac Studio and a Rubin rack are closer than intuition suggests. The rack wins decisively on everything else: NVLink, batching across hundreds of concurrent requests, ECC, multi-tenancy. So a Mac is an excellent single-user inference node, an R&D machine that removes the token meter from experimentation, and a credible appliance for a clinic, a ship or a branch office. It is not a serving platform for a national programme.

The constraints that actually bind

TSMC's advanced packaging is sold out, NVIDIA holds over 60% of 2026 CoWoS output, and HBM production is committed; non-priority buyers face 30-plus week lead times. Compute is rationed, not priced.

Power is tighter still. Singapore hosts about 1.4 GW across 70-plus data centres and allocated 200 MW in August 2026 in four 50 MW tranches, against a PUE target of 1.3. At 190–230 kW a rack, a 50 MW award is roughly 200–250 Rubin racks. National AI capacity is a small, countable number of racks, allocated years in advance. Edge inference — 200–300 W per workstation on existing site power — competes for none of it.

Five decisions, in order

1. Name your sovereignty level per class of data, not per organisation.
2. Secure megawatts before GPUs. Grid and facility readiness is the long pole, measured in years.
3. Buy the platform, not the silicon — and write portability in: open model formats, an OpenAI-compatible serving interface, an internal registry with pinned revisions.
4. Build evaluation capability before infrastructure. A team that can measure whether a 9B model suffices is worth more than an extra rack.
5. Run a portfolio and publish the escalation rate. Frontier for the hard tail, sovereign capacity for regulated volume, local silicon for R&D and the edge.

The organisations that do well over the next three years will not be the ones that picked a side. They will be the ones that measured their workloads honestly and placed each one where it belongs.