The day in AI Daily
AI’s centre of gravity shifted from ever-larger training runs towards the harder business of putting models to work. OpenAI published the first results for Jalapeño, its custom inference chip, claiming 1.5–1.9 times more work per watt at peak throughput and 1.7–3.6 times lower end-to-end latency than comparison systems across three large open models. The tests used the public InferenceX framework, but OpenAI ran them itself; production qualification is still under way and initial deployment is only planned by year’s end. Even so, working first-party silicon gives OpenAI a credible route to reduce serving costs and latency while it continues using Nvidia and other suppliers for training and inference capacity (OpenAI).
At smaller scales, the same push appeared on desks, factory floors and laboratory benches. Perplexity launched Portable Computer, an agent optimised for Nvidia’s DGX Spark that can run a post-trained 27-billion-parameter Qwen model locally, connect to Gmail, Drive, Slack and GitHub, and switch to cloud models for harder work (Nvidia). IBM released Apache 2.0-licensed Granite 4.2 reasoning models in 3B, 8B and 30B sizes, alongside compact speech-recognition models intended for edge deployment (IBM Research). These are not frontier-capability breakthroughs. Their significance is that useful reasoning, tool use and speech processing are becoming deployable under an organisation’s own operational and data controls.
The deeper pattern
The day’s common thread was vertical integration: AI companies are trying to control more of the path between a model’s reasoning and its effects in the world. That path now includes chips, local computers, enterprise connectors, robot controllers, scientific instruments and the institutions that decide which synthetic outputs count.
Jalapeño is the clearest infrastructure example. Training determines what a model can potentially do; inference determines how often, how quickly and at what cost it can do it for users. This matters disproportionately for agents because a task may require dozens of sequential model calls. Even modest latency at each step compounds. OpenAI’s benchmark suggests specialised hardware can improve both responsiveness and power efficiency, but the comparison is not yet an independent verdict: it uses selected models and operating points, normalises power using published ratings, and does not establish reliability or total cost at fleet scale (OpenAI).
Portable Computer takes the same economics to the individual machine. Local execution offers privacy, predictable availability and effectively zero per-token charges after the hardware is purchased. Yet connecting an always-on agent to email, files and workplace messaging also expands its authority and attack surface. “Local” protects data from routine cloud processing; it does not, by itself, make an agent safe from malicious documents, compromised connectors or mistaken actions. The important test will be whether local-first products expose permissions, logs and confirmation boundaries as clearly as they advertise privacy.
Google Cloud is approaching the control problem from the opposite direction: not moving computation onto a user’s desk, but packaging it inside governed industry systems. Gemini Enterprise for Financial Services entered preview with a managed research agent, more than 50 specialised skills and 13 connectors, while the legal edition targets work including contract review, diligence, regulatory monitoring and privacy requests. Google says these systems inherit permissions, retain audit trails and ground outputs in attributable sources (financial services, legal). Those controls are more consequential than another chatbot interface, but they remain provider claims from preview-stage products. Accuracy, privilege preservation and liability under real workloads have not yet been demonstrated publicly.
Physical AI extends the integration problem beyond information systems. Nvidia announced the Jetson Orin Nano 2, rated at 78 TOPS with 8GB of memory and an eight-core CPU; it claims twice its predecessor’s compute and 40 per cent lower power use in a 15-watt mode, with availability planned for the first half of 2027 (Nvidia). Meanwhile, Generalist reportedly raised about US400 million round, to develop models rather than robot hardware (Axios). Capital is betting that a reusable “robot brain” can become a valuable layer across many machines. That thesis still depends on proving generalisation outside carefully prepared demonstrations, where errors carry physical costs.
Two scientific developments show both the promise and the evidentiary gap. NTT reported an autonomous thin-film deposition loop combining robotics, optical measurement and Bayesian optimisation. It says the system ran experiments about three times faster, produced a single-crystal beta-gallium-oxide film and extracted a human-readable process rule whose usefulness was tested through re-optimisation; the underlying research was published in Nature Communications (NTT). By contrast, Accelerated Understanding’s non-Transformer neural-operator system is presently a much larger claim: the company says it has trained models up to one trillion parameters and tested inference contexts exceeding five trillion physical data points, but has not supplied independent benchmarks establishing accuracy, cost or superiority on the proposed chip-design, weather and robotics applications (Accelerated Understanding).
Creative industries are building their own control layers. Stability AI’s US$76 million Series B includes Sony Music, Universal Music, Warner Music and Electronic Arts, tying a model company more closely to rights holders and production channels (Stability AI). At the same time, ARIA will exclude wholly AI-generated tracks from its charts and awards while retaining eligibility for substantially human-made, AI-assisted recordings (Associated Press). One development draws incumbents into AI production; the other establishes a boundary around recognition. Together they suggest that synthetic media’s next phase will be shaped as much by licences, provenance and distribution rules as by model quality.
What to watch next
-
Whether OpenAI publishes independently reproducible Jalapeño results—or fleet-level figures for utilisation, failure rates, energy use and cost per completed task—when deployments begin by the end of 2026.
-
Whether Google’s legal and financial previews produce named customer evidence showing lower end-to-end error rates, preserved access controls and successful audits, rather than only faster document production.
-
Whether physical-AI claims survive external testing: Accelerated Understanding should release task definitions and comparative accuracy results, while Generalist should demonstrate one model transferring across materially different robots and unprepared environments without extensive retraining.
Editorial note
Most of today’s technical evidence came from vendors, and several products are in preview, pre-production or scheduled for 2027. NTT’s peer-reviewed result is the strongest scientific anchor; Jalapeño’s measurements are detailed but self-run, while the five-trillion-context claim lacks independent validation. The supplied signal set also contained no substantial new AI-security incident or evaluation, so this edition may underrepresent risks emerging outside public product announcements.