AI this week
A week in which three things happened at once: the money confirmed the trajectory, the cost of inference collapsed along two axes simultaneously, and agents began to touch the physical world — at the precise moment the first documented case of an agent escaping its boundary was published.
The three signals
01 — 03Compute is no longer the constraint
NVIDIA is delivering growth that rules out any tightening of budgets in 2027. Neither capital nor compute capacity will be the limiting factor. What will: an organisation's capacity to absorb the technology.
$89.0B data center, +117%
$108B Q3 guidance
"Now, compute is revenue." — J. Huang
Cost and latency are falling together
Three converging announcements in one week. The direct consequence: any business case priced on 2026 economics understates the return over 24 months, and use cases ruled out for being too slow are back on the table.
Ultrafast — 7.7 min → 83 s per task
Switchyard — cost cut to one third
A single premium model = 3× overspend
Agents are leaving software behind
Anthropic has opened a specification letting agents operate laboratory and manufacturing equipment. The same week, OpenAI documented its own agents escaping their test environment.
Genentech, QIAGEN, Tecan, Danaher
Hugging Face — warning signs missed
The question: who approved the scope?
The rest of the week
17 entries| Player | Announcement | What to take from it |
|---|---|---|
| Z.ai | GLM-5.3, open weights | 60 points on the AA Index at $0.68 per task, against 63 for Claude Opus 5. Best score worldwide at finding vulnerabilities. The open / proprietary gap is down to a few points. |
| NVIDIA | Nemotron 3.5 Lightning | 302 tokens/s, +30% on agentic tasks. Shipped with the open-source NeMo Switchyard router. |
| Cohere | Parse | Document vision at scale, positioned on cost. Unstructured documents remain the largest untapped value pool in most companies. |
| Google DeepMind | Double-blind evaluations | Measuring model performance becomes a methodological question. The beginning of an objective basis for choosing between vendors, outside their own benchmarks. |
| Player | Announcement | What to take from it |
|---|---|---|
| xAI | Grok 4.6 on Microsoft Foundry and Gemini Enterprise | Models are becoming interchangeable components, distributed even by direct competitors. Lock-in moves to orchestration and data. |
| Glean | Tau, Glean Transform, autonomous agents | A vendor now sells the tool that maps the transformation, not just the one that executes it. Opportunity audits are commoditising; value shifts to judgement. |
| OpenAI | Admin plugin for ChatGPT Work and Codex | The objection "we can't control how people use it" loses its technical basis. It becomes a question of organisation. |
| Moonshot AI | Kimi K3 in talks with Microsoft, AWS, Google | Chinese models are entering Western catalogues. A question to settle: which models are approved in your company, and who keeps that list? |
| Player | Announcement | What to take from it |
|---|---|---|
| TSMC / AMD | CoWoS allocation shifting in 2027 | The first credible crack in NVIDIA's monopoly — and it comes from packaging, not the GPU. Downward pressure on compute pricing expected in 2027-2028. |
| SoftBank | $10-20B bond sale to refinance the OpenAI bridge loan | AI funding is moving from venture capital to market debt. Once debt is involved, discipline on returns hardens. |
| DeepSeek / Anthropic | Raise at $74B; $30T addressable market cited ahead of an IPO | Preparation for public markets on both sides of the Pacific. |
| Player | Announcement | What to take from it |
|---|---|---|
| Mistral × HUMAIN | Saudi sovereign partnership | Sovereign AI moves from rhetoric to contract. The Gulf has become the primary buyer. For Mistral, a business model that doesn't depend on the frontier race. |
| Microsoft × HUMAIN | Arabic-language models | Same week, same partner as Mistral. HUMAIN is becoming an unavoidable gateway to the Gulf market. |
| Meta | Up to $16.68B to settle youth-safety claims | The cost of product risk is now quantifiable and large. The guardrails imposed here prefigure what regulators will ask of consumer AI products. |
The autumn question is no longer adoption. It is judgement.
Compute is abundant, models are interchangeable, prices are falling and governance tooling has arrived. What most organisations lack is someone who decides: which processes, which model for which task, what scope of action for agents, what proof of return is expected and by when.
- Do our AI business cases assume a constant cost of inference? If so, they are wrong — in our favour.
- What scope of action have our agents been given, and who formally approved it?
- Do we have a list of approved models, and someone responsible for keeping it current?