China's Unitree surged 460% on its Shanghai debut, taking the humanoid robot maker to a $50 billion valuation in one trading session. AT&T cut its AI coding costs by 56% just by routing easy questions to cheaper models instead of its most expensive one. Washington is reportedly telling 35 countries they can't run US and Chinese AI systems side by side. One story is a market betting big on physical AI, one is a business quietly working out how not to overpay for it, and one is the map underneath both getting redrawn.

Summary: Unitree Robotics, the Hangzhou maker of backflipping humanoid and quadruped robots, jumped as much as 629% on its 19 August Shanghai debut before settling near a 460% gain, taking its valuation from a $904 million IPO to roughly $50-66 billion in a single session. That put it ahead of established Chinese tech names like Baidu and JD.com, on an offering retail investors oversubscribed more than 8,000 times. DeepSeek was among the investors, putting AI's most prominent language-model maker directly into a robotics stock for the first time.

Why it matters: A robot maker briefly outvaluing two established tech giants in one trading day is the kind of number that gets attached to a whole category, not just one company. Unitree ships real, working machines, but an 8,000x oversubscription rate says as much about how much capital is chasing "physical AI" right now as it does about the robots themselves. The last time enthusiasm moved this fast in AI, the repricing that followed wasn't gentle.

What to do:

Source: Bloomberg | Fortune

Summary: AT&T used LiteLLM, an open-source model router, to send simple employee AI requests, like document and code summaries, to cheaper open-weight models such as Meta's Llama and Google's Gemma, instead of paying frontier rates from Anthropic or OpenAI for every task. The company cut coding-related AI costs by up to 56%, with only a 2% dip in output quality, and now aims to route 60-70% of employee queries through open models, up from 40% today.

Why it matters: The default for two years has been one premium model for everything, on the assumption that quality only moves one way. AT&T's numbers say most of what a workforce actually asks AI to do doesn't need the most expensive model available, and a 2% quality trade-off for a 56% cost cut is not a close call for most budgets. Expect this to become the standard playbook once more companies see numbers this clean.

What to do:

Source: PYMNTS | Tech Startups

Summary: The US State Department is reportedly preparing to tell 35 countries already signed on to its Pax Silica AI framework that they can't also join a rival Chinese AI initiative, according to an internal draft reported in mid-August. China responded on 19 August by urging countries to defend their "digital sovereignty" and reject what it called zero-sum thinking. Neither government has confirmed a final policy, but the draft marks the first explicit either-or ultimatum either side has put in writing.

Why it matters: Every business selling AI-powered software across borders has taken it for granted that the tools underneath keep working the same way everywhere. A formal ultimatum turns that assumption into a live question. Countries forced to pick a lane don't just choose a diplomatic camp, they choose which models, chips, and cloud providers their regulators will actually allow.

What to do:

Source: The Next Web | China Global South

Summary: An AI-manipulated clip showing a costumed Darth Vader addressing a San Diego city council meeting to mock Flock Safety's surveillance cameras circulated on 20 August, and CNN, Fox News, UPI, The Hill and dozens of other outlets ran it as real news. San Diego's own public meeting archive shows something much shorter: a committee member calls for a speaker named Darth Vader, nobody appears, and the meeting moves on. The fabricated chyron, audience reactions, and Star Wars jokes were all AI-generated.

Why it's interesting: This wasn't a small blog getting fooled. Multiple national newsrooms with fact-checking desks ran a talking Sith Lord as legitimate public comment before anyone checked the source footage.

Key takeaway: If newsrooms with editors and standards desks can run a fake Darth Vader as fact, checking your own AI-sourced clips before you share them isn't optional anymore.

Source: Techlicious | Washington Times

Summary: An American Express customer testing a support chatbot named "David" over a billing issue asked it to write a haiku and answer a string of odd questions, including a Monty Python reference, after it claimed twice to be a real person. The exchange, posted to X on 19 August, ends with the bot admitting it's AI once the questions got specific enough that a scripted human persona couldn't keep up.

Why it's interesting: A financial services chatbot claiming to be human, on the record, is exactly the kind of screenshot that turns into a regulatory conversation.

Key takeaway: If your business runs a support bot, a customer testing whether it will admit what it is should never be the first time you find out the answer.

Source: Trung Phan on X | The Poke

Grok Bot - xAI's app for building named, always-on AI agents that handle real tasks like research or email triage, connected to Gmail, Slack, and Calendar, and set to run on a schedule. Best use case: a standing team member for repetitive work nobody wants to own. x.ai/news/introducing-grok-bot

Unsloth Studio - an open-source, no-code interface for fine-tuning open models like Llama or DeepSeek-R1 locally, on a single consumer GPU, without writing training code. Best use case: teaching a model your own specialised knowledge without hiring an ML engineer. unsloth.ai

LiteLLM - the open-source model router AT&T used to cut its AI costs, giving one interface to more than 100 providers with automatic routing to the cheapest model that can handle a given task. Best use case: testing whether your own AI spend has the same easy savings AT&T found. litellm.ai

ElevenLabs Conversational AI - build voice agents that answer calls, book appointments, and handle routine questions in a natural voice, with sub-300ms response times across dozens of languages. Best use case: a phone line that doesn't put customers on hold. elevenlabs.io

Claude Academy - Anthropic's free learning hub, with 355 tutorials and prompting guides plus structured courses on Claude, Claude Code, and the API. Best use case: getting a team genuinely competent with the AI tools it already pays for. academy.claude.com

Context: AT&T just proved that routing simple AI tasks to cheaper models can cut costs by more than half with almost no quality loss. Most businesses have never mapped which of their AI tasks genuinely need a frontier model and which don't.

Prompt: "You are a cost analyst reviewing my AI usage. Here is a list of tasks my team asks AI to do in a typical week: [list each task - what it's for, how often it happens, and which AI tool or model currently handles it]. For each task: (1) rate whether it genuinely needs a frontier model's reasoning or could be handled by a cheaper open-weight model, (2) flag any task where a wrong answer would be costly, since those should stay on the expensive model regardless of price, (3) estimate what percentage of my total AI tasks could plausibly move to a cheaper model without hurting the actual outcome, (4) suggest a router or workflow for making that switch practical."

Do not: Do not assume a cheaper model is automatically worse. AT&T's own testing found only a 2% quality dip after moving over half its coding tasks to open models.

If / then: If more than a third of your AI tasks could plausibly move to a cheaper model, that's a cost review worth scheduling this month, not a someday project.

Example use case: A ten-person marketing agency ran this over its weekly AI tasks: blog drafts, ad copy variations, client reporting summaries, and campaign strategy. Ad copy variations and reporting summaries moved to a cheaper model with no client noticing. Blog drafts and campaign strategy stayed on the frontier model. Their monthly AI bill dropped by about a third.

Summary: Nvidia CEO Jensen Huang met with Rebellions co-founder Sunghyun Park at Nvidia's Santa Clara headquarters this week to discuss a possible investment, partnership, or acquisition, Bloomberg reported on 21 August. Rebellions, a Korean AI inference-chip startup last valued at $2.3 billion, has raised roughly $850 million from SK Hynix, Samsung Ventures, and Arm, and its CFO said days earlier the company is also preparing its own IPO.

Why it's important: Nvidia already dominates the training-chip market it sells into its own biggest customers. Talking to a company that specialises in inference, the cheaper, higher-volume workload that actually runs AI day to day, would extend that reach into a part of the chip market Nvidia doesn't already own outright.

Business takeaway: If your AI costs are dominated by running models rather than training them, watch whether this deal closes. It would put one more piece of the inference chip supply chain under a single supplier.

Summary: OpenAI dropped GPT-5.6 Sol's API price by more than 20% on 21 August, taking input costs from $5 to $4 per million tokens and output from $30 to $20, through at least 21 November. It's the second cut to the Sol family in under a month, and it lands squarely under Claude Opus 5's pricing.

Why it's important: Last week this newsletter flagged that promotional AI pricing tends to carry an expiry date once it appears. This is the other half of that pattern: a lab cutting its own frontier price twice in a month to stay competitive, with a hard end date already attached.

Business takeaway: If Sol is part of your stack, use the discount while it lasts, but budget around the 21 November date it's set to end, not the current rate.

Summary: Google's Agent2Agent (A2A) protocol formally joined the Linux Foundation's Agentic AI Foundation on 20 August, placing it under the same neutral governance structure as Anthropic's Model Context Protocol (MCP). Both now sit inside a vendor-independent foundation instead of under either company's direct control.

Why it's important: Two competing labs putting their agent-communication standards under the same outside governance runs against the usual AI story, where each vendor tries to make its own format the default. A neutral home for both makes it more likely agents built on different platforms can actually talk to each other.

Business takeaway: If you're choosing between building on A2A or MCP, neither pick looks like the safer long-term bet anymore. Both now answer to the same outside foundation instead of a single company's roadmap.

That's it for this week's Weekly Round-up. Forward this to one person still paying frontier prices for AI tasks that don't need them. See you next week. - Nicolas