FORSMILE
Issue #1Published August 26, 2026 / Covering Aug 20-26日本語で読む

7 AI stories this week — OpenAI publishes measured results for its own inference chip, Porsche sells its IT consultancy to TCS

The week in one place

The theme this week was owning the stack. OpenAI published the first measured results for Jalapeño, its own inference chip, and used the occasion to argue for designing models, serving software, silicon, memory and networking as one system. Every figure is OpenAI's own measurement, and the comparison systems are described only as "leading commercially available AI systems" — worth discounting accordingly. On the tooling side, agents kept moving to where people already are. GitHub Copilot can now run as a shared agent session inside a Microsoft Teams thread, and the Copilot app's extension surface — MCP servers, plugins, skills and canvases — reached general availability in a single tab. On the corporate side, Porsche signed an agreement to sell its management and IT consultancy MHP to TCS. Note that this is a signature, not a completion: regulatory approval is still outstanding. In safety and policy, Anthropic launched a $5 million independent research grant programme and OpenAI started a blog for a new team working on concentration-of-power risk.

Market & Industry

OpenAI publishes the first measured results for Jalapeño, its own inference chip

1.5-1.9x more work per watt and 1.7-3.6x lower end-to-end latency, as measured by OpenAI itself.

On August 25, 2026, OpenAI published measured results for Jalapeño, its first custom inference chip. Testing on InferenceX, a public benchmark from SemiAnalysis, across GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T, OpenAI reports 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the comparison systems. For highly interactive workloads it reports 2.1 to 4.1 times higher performance. On Kimi K2.5 1T, the largest public model tested, it reports roughly 1.5 times higher peak performance per watt and 3.4 times lower end-to-end latency. The chip is rated at 700 watts, with measured sustained power at or below 550 watts on the workloads tested. All figures are OpenAI's own measurements. The comparison systems are described only as "leading commercially available AI systems" and are not named. Results were normalised using each accelerator's published chip power rating.

So What

Downward pressure on inference cost is now arriving from a route that does not run through GPU procurement. But this is vendor-run measurement rather than third-party verification, so treat it as a signal of intent until it shows up in pricing.

Funding & Corporate

Porsche signs agreement to sell its IT and management consultancy MHP to TCS

Signed on August 24 — but not closed, and Porsche's release states no deal value.

On August 24, 2026, Porsche announced it had signed an agreement to sell the management and IT consultancy MHP to Tata Consultancy Services. Porsche frames it as a milestone in its "Sportwagenschmiede 35" strategy of concentrating on its core business. MHP employs more than 4,500 people and is headquartered in Ludwigsburg, near Stuttgart. It will keep its brand and continue to operate as an independent consultancy. Porsche, TCS and MHP also plan a partnership in digital solutions and artificial intelligence. Porsche's release states explicitly that the acquisition remains subject to customary regulatory and competition law approvals and is expected to be finalised in the coming months, so **the transaction is not complete**. The release also gives no deal value, and no figure is stated here that the parties themselves have not published.

So What

A carmaker letting go of its in-house IT consultancy and sourcing AI capability through a partner instead. It is evidence that large enterprises are actively choosing options other than building AI teams internally.

Developer Tools

The GPT-5.6 family becomes available in AWS Kiro

Sol, Terra and Luna land in Kiro, with a reported ~82% cost reduction per successful task on Terminal-Bench 2.1.

On August 24, 2026, OpenAI announced that the GPT-5.6 model family — Sol, Terra and Luna — is available in Kiro, the AWS software development agent. Kiro turns high-level intent into requirements, technical designs and executable tasks, an approach OpenAI describes as spec-driven development. OpenAI and AWS jointly optimised the Kiro environment and the models. Their testing found that on Terminal-Bench 2.1, GPT-5.6 Terra completed successful tasks at roughly 82% cost reduction. The announcement does **not state what that 82% is measured against** — an earlier model, or the same model outside Kiro. It also mentions checking implementation correctness with property-based testing.

So What

An example of why a model name alone does not tell you what an agent will cost. But because the 82% has no stated baseline, it is not evidence that you would see the same reduction — treat it as a reminder to check the comparison conditions behind any vendor cost figure before planning around it.

Developer Tools

GitHub Copilot runs shared agent sessions inside Microsoft Teams

Mention @GitHub in a Teams thread to start a cloud agent session everyone in the conversation can watch and steer.

On August 21, 2026, GitHub shipped the ability to start a GitHub Copilot cloud agent session by mentioning @GitHub in a Microsoft Teams channel, thread or direct message. Anyone in the conversation can ask questions, add context and help direct the work, while only participants with write access to the repository can trigger Copilot to make changes. The task runs asynchronously in a secure cloud sandbox, progress is followed in the channel thread, and the resulting artifacts can be picked up from a terminal, the GitHub Copilot app or any IDE. It is in **public preview** and available with paid GitHub Copilot plans.

So What

Agent execution moves from the IDE into the meeting, and authority is decided by repository write access rather than by who asked. That is a reason to revisit permissions on the chat side, not just in the repo.

Developer Tools

The Customize tab in the GitHub Copilot app reaches general availability

MCP servers, plugins, skills and canvases are now browsable from a single tab.

On August 25, 2026, GitHub made the Customize tab in the GitHub Copilot app generally available. It brings four kinds of extension — MCP servers, plugins, skills and canvases — into one place, with a featured view, browsing by type, and category browsing plus trending options for MCP servers. One highlighted example is a canvas that delegates Azure DevOps backlog work to Copilot: triaging issues, prioritising backlogs, assigning follow-ups, and handing work over to investigate, implement or prepare for review. The featured view is positioned as an entry point for people who do not yet know which extension they need.

So What

Extension discovery is now a product feature rather than a documentation problem. If you publish your own skills or MCP servers, the path to being found shifts from "get people to read the README" to "appear in a listing".

Regulation & Safety

Anthropic puts $5 million into independent research on AI and user wellbeing

Grants provide funding, model access and technical support to researchers building open-source evaluations.

On August 25, 2026, Anthropic announced a $5 million grant programme funding independent research into how AI affects users' wellbeing. Grantees receive direct funding, access to Anthropic's models and technical support, work fully independently, and publish their results as open-source evaluations any developer can use. The announcement frames wellbeing as hard to evaluate precisely because a single response cannot be judged in isolation. It gives two examples: a user in distress may not disclose thoughts of self-harm immediately, so the need for a more cautious response only becomes clear over a long conversation; and standard diet and exercise advice to someone asking about weight loss could be actively harmful for a user with a history of disordered eating. Anthropic invites clinicians, psychologists and methodologists to take part.

So What

The industry is moving toward evaluating harm across a conversation rather than per response. If you build anything with long-running dialogue, this is where shared evaluation baselines are likely to come from.

Regulation & Safety

OpenAI launches AI Futures, a blog for its new team on concentration-of-power risk

The Strategic Futures team asks how a free society should be restructured to accommodate transformative AI.

On August 20, 2026, OpenAI launched AI Futures, the blog of its new Strategic Futures team, with a first post by Dean Ball. The team's stated question is how a free society should be restructured to preserve individual rights and agency while accommodating the emergence of transformative AI — what the AI safety and policy community calls concentration of power risks. The team argues this category is, in the long run, the largest and most conceptually difficult risk to address. The opening post builds on the observation that political and military power has historically rested on human labour and on large-scale consent. **The blog states explicitly that posts reflect the views of their authors and not organisational positions of OpenAI.**

So What

Policy attention is widening from "is this model dangerous" to "what does this do to power structures". If that framing takes hold, the shape of obligations placed on operators changes with it, so it is worth tracking early.

Watching (no confirmed primary source yet)

  • EU AI Act transparency obligations became enforceable on August 2, 2026. That falls outside this week, so it is not an item here; a concrete enforcement action would be.
  • OpenAI's "Offering Zero Data Retention for frontier models" is listed as August 20 on the news index, but the article page itself is dated August 19 — outside this issue's range. Recorded here rather than carried as an item.
  • Reports around August 22 about Anthropic moving into hardware are secondary only; no primary announcement from the company was found, so it was left out.
  • Reports of a chipmaker considering a stake in Perplexity, and the valuation attached to it, have no confirmation from the parties involved and are not covered.

"Primary" links go to the announcing party's own publication (company blog, press release, official docs). "Reporting" links go to news coverage or third-party analysis. Figures and dates are as verified on the publication date.

← All Weekly AI News issues