AI Digest — 30 August 2026

The operational layer of AI is repricing. The hardware fund closes, the original crowdsourced-labour marketplace shuts down, adoption benchmarks confirm agents are now load-bearing infrastructure, and OpenAI continues its quiet product consolidation. Each story points in the same direction: the decisions that looked optional twelve months ago are now consequential.

Andreessen Horowitz Closes a $1.1 Billion Fund Dedicated to AI Hardware

a16z announced on 28 August that it has closed the Machine Age Fund, a $1.1 billion vehicle it describes as hardware-first — the firm's first fund dedicated exclusively to the physical layer of AI. General partners Raghu Raghuram and Martin Casado will lead investments across chips, memory, networking, storage, data centres, robotics, and home appliances.

The fund's investment thesis rests on a supply-side reading of the moment: hardware has grown from a small fraction of a16z's deal flow to more than 20 per cent, and the bottleneck now constraining what AI can do in production is physical rather than algorithmic. Power capacity, memory bandwidth, and interconnect speed are what limit frontier inference at scale. The fund formalises a strategy that has already included Skydio, Anduril, Waymo, and more recent bets such as Unconventional AI, Nexthop, and Heron Power.

For operators, the practical read is this: a concentrated VC focus on the physical stack signals that scarcity in compute, power, and cooling is real and that new entrants are coming. Teams that have built plans assuming commodity AI compute will remain cheap and plentiful should model the alternative. Infrastructure-as-a-service pricing has moved materially in 2026; the Machine Age Fund suggests that trajectory continues rather than corrects.

Amazon Closes Mechanical Turk After 21 Years

Amazon announced on 25 August that Mechanical Turk will close on 30 September 2026. Jeff Bezos described it at launch in 2005 as "artificial artificial intelligence" — a marketplace where businesses posted small digital tasks and humans completed them at scale. At peak, more than 500,000 workers participated. SageMaker Ground Truth, Amazon's managed labelling service built on Mechanical Turk, closes at the same date.

The replacement is not a single successor. Data labelling and annotation have migrated to a newer generation of AI-native platforms — Scale AI, Mercor, and Prolific — which use AI to improve the consistency and throughput of human annotators and, in some task categories, replace them entirely. Amazon stopped accepting new Mechanical Turk customers in July, signalling the wind-down months before the formal announcement.

The symbolic weight is hard to overstate. Mechanical Turk was the infrastructure that made early AI progress economically feasible: it labelled the datasets that trained the models that are now displacing it. Any operator still routing data-labelling work through Mechanical Turk needs an alternative pipeline in place before 30 September. Any operator watching the broader labour-displacement question has a concrete data point: the platform that first industrialised human-AI task division is now obsolete.

Temporal's Report: 80.8 Per Cent of Engineers Use AI Agents Daily

Temporal published its 2026 State of Development Report this week, drawing on a survey of 554 engineers and engineering leaders in the US and UK. The headline: 80.8 per cent report using AI agents daily or more frequently, up from 47.3 per cent a year ago — a 70.8 per cent relative increase in twelve months.

The adoption curve tells part of the story; the reliability gap tells the rest. The teams that moved fastest on agents are now the ones confronting failures in observability, auditability, and security. Agent use has gone from experimental to load-bearing for most respondent organisations, but governance infrastructure has not kept pace with the deployment rate. The report's second headline, underneath the adoption figure, is that teams are running agents they cannot fully inspect or control.

• 80.8% of engineers use AI agents daily, up from 47.3% a year earlier
• Survey base: 554 engineers and engineering leaders, US and UK
• The gap opening up is not adoption — it is observability and governance at production scale

For an operator assessing where their team stands: if daily agent use is below 80 per cent, your organisation is trailing the median engineering peer. If it is at or above that level, the Temporal data suggest the priority is now governance infrastructure, not further adoption. The question is not whether agents are being used — it is whether anyone can explain what they did when something goes wrong.

OpenAI Retires the DALL-E GPT in ChatGPT Today

The dedicated DALL-E GPT plugin inside ChatGPT reaches end-of-life today, 30 August, as reported by Tom's Guide and NotebookCheck. Users who generated images through that tool should download them now; storage beyond today is not guaranteed. User-created GPTs that have image generation enabled are unaffected.

Image generation inside ChatGPT is not going away. ChatGPT Images 2.0 covers the same core functionality and is available on all plan tiers, including the free plan. OpenAI had already retired the standalone DALL-E API for developers earlier in 2026. The retirement of the GPT plugin is the final step in consolidating image generation into the main ChatGPT product surface.

Operators who have image-generation workflows that specifically invoke the DALL-E GPT should redirect them to ChatGPT Images or the API-level image endpoints. The broader pattern — a legacy plugin replaced by an integrated product feature — is consistent with what OpenAI has done across Atlas, o3, and several other tools in 2026. The product surface is narrowing as each function matures into the core.

The week's pattern is not coincidental. Hardware, labour, tooling, and adoption metrics are all signalling the same structural shift: the AI stack is becoming critical infrastructure, with the costs and governance obligations that implies. Operators who still treat AI deployment as an experiment will find the pricing, the workforce transitions, and the governance expectations have moved on without them.