Million Tokens on a MacBook

Six months ago, running a serious AI model required data centre hardware. Most procurement and strategy teams have not caught up with what happened next.

Since then, the floor has dropped every month: a $600 consumer GPU ran 70-billion-parameter models in February, Google halved inference memory requirements and a full voice assistant ran locally on Mac in March, and by April frontier models were practical on everyday MacBooks and running natively on iPhones. Now ds4 extends the trend to a million-token context window — roughly equivalent to an entire codebase or a shelf of documents in a single query — on a laptop.

Salvatore Sanfilippo — known as antirez, creator of the database engine Redis — built ds4 to run one specific model, DeepSeek V4 Flash, on Apple's hardware. It requires a MacBook with 128GB of memory and uses aggressive compression to fit what would normally demand cloud infrastructure onto a laptop. It builds on foundational open-source work and follows DeepSeek V4's launch two weeks earlier, which challenged the pricing assumptions of every major AI provider.

It is a narrow tool — one model, one hardware platform — but it is a proof point for what is now possible without a cloud contract.

This technical compression is landing against a stark economic backdrop. Big Tech's combined AI infrastructure spend hit $725 billion this year, the Financial Times reported on the eighth of May, pushing Amazon, Google, Microsoft, and Meta's collective free cash flow to a decade low. Downstream, the cost pressure is already visible: "tokenmaxxing" — engineers maximising their use of AI coding assistants and chat tools — has driven individual monthly bills as high as $150,000 at some firms, according to the Times of India.

When infrastructure spending runs this far ahead of revenue, the pressure to recoup costs shifts to customers. That means tighter free tiers, more aggressive bundling, or straight price increases on the APIs that many products now depend on.

Running AI locally is no longer just about data privacy. For any team paying serious cloud AI bills, it is becoming a straightforward cost decision — and ds4 means that calculation now extends to the kind of long-context work that, until this week, had no local option at all.

The trend is real, but the limits matter. A 128GB MacBook is a £3,500+ machine — most teams issue laptops with a quarter of that memory. The compression ds4 uses to fit the model involves accuracy trade-offs that may not be acceptable for every use case. And it runs exactly one model, not a menu. For teams that need flexibility across multiple AI providers or run standard-spec hardware, cloud APIs remain the practical default.

The thing that would reverse this trajectory: a next wave of AI models so large or complex that they push back beyond what consumer hardware can handle. That has not happened yet.

---

Several of this week's other stories reinforce the same cost-and-control tension that runs through the editorial above.

• Sakana AI released RL Conductor, a 7B parameter model that learned to route queries across GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro — sending easy tasks to cheaper models and hard tasks to expensive ones. Sakana claims 83% token cost reduction over static routing. Another approach to the same problem ds4 tackles from the hardware side: making AI cheaper to run.

• OpenAI launched three real-time voice models for reasoning, translation, and transcription via its Realtime API, replacing the need to stitch separate speech-to-text, language model, and text-to-speech services together. This is the cloud side of the argument — capabilities that are hard to replicate locally, designed to keep developers on the platform.

• Cloudflare cut 1,100 employees — 20% of its workforce — citing a 600% increase in internal AI usage as the direct driver. It is the most explicit AI-to-headcount attribution from a public company this year, and a reminder that the cost pressure AI creates runs in both directions: infrastructure bills go up, headcount goes down.

• Anthropic, OpenAI, and Google began embedding engineers directly inside enterprise customers, moving from model vendors to implementation partners. HCLTech reported 2-3% AI-driven revenue deflation as a result — the clearest sign yet that AI providers are competing not just for API calls, but for the services revenue around them.

---