A company best known for robots that dance on Chinese New Year television just became worth more than most of the aerospace industry. Unitree Robotics listed on Shanghai's STAR Market on August 19, 2026, priced at 150.8 yuan a share, and opened at 1,100 yuan. That is a 629 percent pop, a market capitalisation near 445 billion yuan or roughly $66 billion at the peak, and a retail order book oversubscribed 5,500 times. It closed at 845 yuan, up 460 percent, worth about $50 billion. Meituan's 8.7 percent stake returned roughly 70 times.
The rest of the day belonged to the people supplying the picks and shovels. Google secured a warrant to buy up to $12.2 billion of Marvell stock tied to custom chip purchases, Nvidia is weighing an investment in Mercor at a $20 billion valuation, Samsung raised foundry prices by up to 15 percent, and Fractile reached $6.5 billion off an Anthropic inference chip order. OpenAI's CFO meanwhile told an all-hands that the company will be public in 2027. Here are the 18 stories that matter for August 20, 2026. For running coverage of every release this month, bookmark our AI industry news and trends hub.
1. How Much Did Unitree Robotics Stock Rise on Its Debut?
Unitree Robotics shares opened at 1,100 yuan against an IPO price of 150.8 yuan, a gain of 629 percent, and closed the first session at 845 yuan for a 460 percent rise. The Hangzhou company raised 6.1 billion yuan, about $904 million, and became the first humanoid robot maker listed in mainland China. Retail demand oversubscribed the offering roughly 5,500 times. The peak intraday valuation reached about 445 billion yuan, near $66 billion, settling around $50 billion at the close.
Two numbers give that context. The average first-day pop for new Chinese listings this year has been about 279 percent, so Unitree more than doubled the norm, and it did so on a day when China's benchmark index fell 3 percent. The company shipped more than 5,000 humanoid units in 2025, which is real volume by the standards of a sector where most competitors count deployments in the dozens. Meituan, an early backer with an 8.7 percent stake, saw roughly a 70x return on paper.
My take: this is the moment humanoid robotics stopped being a research story and became a public-market asset class, and the price reflects scarcity as much as fundamentals. There is exactly one pure-play listed humanoid maker in China and 5,500x retail oversubscription tells you what happens when enormous demand meets a small float. I would not read the peak $66 billion as a considered valuation. I would read the 5,000 units shipped as the number that has to keep growing, because that is the only thing underneath the price.
2. Why Unitree's Valuation Reprices Every Robotics Company
Unitree's debut immediately reset comparable valuations across the robotics sector, making US competitors look cheap on a relative basis and giving every private humanoid startup a public marker to point at in their next fundraise. That repricing is the second-order effect that will matter longer than the first-day chart.
Robotics has been the hardest AI subsector to value because almost nothing traded publicly. Investors had funding rounds and press releases, with no market-clearing price. A listed pure-play changes that overnight. It also arrives on top of the physical AI funding data we covered yesterday, where the sector raised $47.4 billion across 521 deals in the first half of 2026, up around 80 percent year over year. Capital was already rotating into machines that act in the world. Now there is a public benchmark to mark it against.
My take: expect a wave of humanoid fundraising announcements over the next six weeks, priced off this comparable, and expect most of them to be aggressive. The honest caveat is that Unitree's premium reflects a Chinese retail market with limited float and a domestic manufacturing base that makes its unit costs genuinely hard to match. Reading it straight across to a US startup with no shipped product is the mistake people will make. See yesterday's roundup on physical AI funding for the underlying capital picture.
3. Google Gets a $12.2 Billion Warrant on Marvell Stock
Marvell Technology granted Google a warrant to purchase up to 58.97 million shares at an exercise price of $206.58, worth as much as $12.2 billion if fully exercised, exercisable until August 18, 2033. It accompanies a commercial agreement signed July 29 covering AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute built to work with Google's TPU ecosystem. Marvell stock rose more than 10 percent, Broadcom fell around 3 percent.
The vesting structure is the interesting part. Roughly 1.4 million shares vest in year one, and the rest unlock in tranches tied to every $500 million of cumulative chip purchases Google makes. Google's ownership therefore scales in direct proportion to how much it buys, which converts a supplier relationship into an equity alignment without Google paying anything up front. A fully exercised position would make Google Marvell's fifth-largest shareholder. Broadcom remains Google's primary custom chip partner under a separate agreement running through 2031, which explains its share reaction.
My take: this is a smarter deal structure than a straight investment and I expect it to be copied. Google gets supply security and upside without capital outlay, Marvell gets a guaranteed anchor customer and a share price bump, and the vesting schedule makes both sides want the purchase volume to grow. It also signals that Google intends to run a second custom silicon source rather than depending on Broadcom alone, which is a rational hedge given how tight chip supply has become.
4. OpenAI's CFO Commits to a 2027 IPO
OpenAI CFO Sarah Friar told an all-hands meeting that the company will be a public company in 2027, or sooner if the business continues to inflect. Earlier reporting had described Friar pushing for a 2027 listing at around a $1 trillion valuation while Sam Altman preferred 2026. This is the first time the timeline has been stated to staff as a commitment rather than a possibility.
The competitive context matters. Anthropic has already filed confidentially and is working with Goldman Sachs, JPMorgan Chase, and Morgan Stanley toward a listing that could come this autumn, on the back of more than $11.5 billion in second-quarter revenue and its first positive adjusted operating income. OpenAI reported $6.7 billion for the same quarter with a declining operating margin. If Anthropic lists first and OpenAI lists in 2027, public markets will have priced the smaller-revenue, higher-margin business before they price the larger consumer franchise.
My take: 2027 is the right call for OpenAI and the delay is not a weakness signal, it is a margin signal. You do not want to go public in the quarter your cost of serving is climbing faster than revenue. The line about listing sooner if the business inflects is the tell, because it means the company is watching a specific metric and will move when it turns. Track the roundup on OpenAI's IPO plans for how the valuation talk has developed.
5. Nvidia Weighs a Mercor Investment at $20 Billion
Nvidia is considering an investment in Mercor, the AI data-labeling startup, at a $20 billion valuation, double the $10 billion mark it carried in October 2025. Mercor booked $614 million in gross revenue in the first half of 2026, an annualised run rate near $2 billion. Nvidia has already spent tens of millions with the company in the last quarter alone.
Data labeling was supposed to be the commodity layer of AI, and this valuation says otherwise. The shift is that frontier labs now pay for expert human data rather than bulk annotation, meaning doctors, lawyers, and PhD-level specialists producing reasoning traces and evaluations that models train against. That work does not compress in price the way image tagging did, because the supply of qualified people is the constraint. Nvidia investing rather than only purchasing suggests it sees the data supply chain as strategically scarce.
My take: a $2 billion run rate on human expert data is the strongest evidence yet that the bottleneck in model quality has moved from compute to evaluation. Everyone can buy GPUs. Not everyone can assemble a network of specialists who will sit and produce the reasoning data a frontier model needs. If you are building AI products, the transferable lesson is that your evaluation data is probably worth more than your prompts, and almost nobody budgets for it that way.
6. Fractile Hits $6.5 Billion on an Anthropic Chip Deal
Fractile, an Oxford spinout building SRAM-based inference chips, is raising around $600 million at a $6.5 billion pre-money valuation, up from $1 billion in May 2026 when Accel, Founders Fund, and Factorial led a $220 million round. The jump follows an order from Anthropic worth roughly $250 million. Production-ready silicon is expected in 2027.
SRAM-based inference means keeping model weights in fast on-chip static memory instead of shuttling them from external high bandwidth memory, which removes the memory bandwidth wall that limits how fast a GPU can serve a large model. It is expensive per bit and hard to scale to very large models, which is why the approach has stayed niche. What changed is that HBM supply is now the binding constraint across the industry, so an architecture that needs less of it looks strategically valuable rather than merely clever.
My take: a 6.5x valuation jump in three months on the back of a single $250 million order tells you how badly buyers want an alternative to the current inference stack. Anthropic ordering rather than partnering is the meaningful detail, because it means a frontier lab is willing to bet production inference on non-Nvidia silicon arriving in 2027. That is a real vote. The risk is unchanged: chip startups slip, and 2027 is a promise, not a shipment.
7. Samsung Raises Foundry Prices Up to 15 Percent
Samsung is raising foundry prices by 10 to 15 percent on its SF4 4nm and SF5 5nm nodes and around 10 percent on 8nm, citing AI demand. Its Pyeongtaek SF4 line has run at full capacity since late 2022. Samsung expects AI-related work to exceed 30 percent of foundry revenue.
Foundry price increases propagate slowly and widely. Every chip fabricated on those nodes gets more expensive, including parts that have nothing to do with AI, which means phone modems, automotive controllers, and networking silicon all absorb the increase. Read it alongside the DDR5 memory story from yesterday, where consumer memory prices rose roughly 500 percent as manufacturers redirected wafers to high bandwidth memory. The pattern is the same: AI demand is bidding scarce fab capacity away from everything else, and consumers pay for it downstream.
My take: this is the quiet inflation nobody is charting. Compute cost per token keeps falling because models get more efficient, while the physical inputs underneath the whole industry keep getting more expensive. Those two curves cannot diverge forever. For anyone planning hardware spend into 2027, assume higher unit costs across the board and build that into the budget now rather than discovering it at purchase time.
8. GLM-5.3 Tops the CyberGym Benchmark at 84.5 Percent
Z.ai released GLM-5.3 on August 14, 2026, post-trained on the same 743 billion parameter base as GLM-5.2. It scores 84.5 percent on CyberGym, ahead of both Claude Mythos 5 and GPT-5.6 Sol, and lifts Terminal-Bench 3.0 from 4.6 percent to 28.3 percent. The model is included in every GLM Coding Plan tier at $18, $72, and $160 a month, with no per-token pricing published. Public weights are targeted for around August 28 after further safety testing.
The architectural point is more interesting than the leaderboard position. GLM-5.3 uses the identical base model as its predecessor, so the entire gain comes from reinforcement learning and training environment design. A jump from 4.6 to 28.3 percent on terminal work without touching the base is a strong argument that most labs are leaving large amounts of capability unclaimed in post-training. GLM-5.3 is not a blanket leader, and Fable 5 and GPT-5.6 Sol still post higher numbers on raw CLI coding and general reasoning with tools.
My take: leading a cybersecurity benchmark while holding back the open weights pending safety testing is a notable combination, and the August 28 target is worth marking in your calendar. The broader signal is that Chinese labs have found post-training returns that Western labs have been slower to harvest, which compounds with the H200 shipments now reaching ByteDance and Tencent. Our Kimi K3 review covers where the open-weight leaders currently sit.
9. OpenAI Spends 20 Percent of Inference Compute on Monitoring
OpenAI estimates that monitoring overhead consumes roughly 20 percent of inference compute for GPT-5.6 Sol-class and higher models, and for all Astra model inference. The figure covers reinforcement learning runs and evaluations involving tool use. The company characterises the cost as internal research spend rather than something passed to customers.
One fifth of inference compute spent watching the model rather than serving the user is a striking allocation, and it is the first time a lab has put a number on the safety tax. It also reframes efficiency comparisons. If a frontier lab burns 20 percent of its serving capacity on oversight, then published cost-per-token figures for the most capable tiers understate the true cost of running them responsibly. The fact that it applies to all Astra inference rather than a sampled subset suggests continuous monitoring rather than spot checks.
My take: I want more labs to publish this number, because it is the clearest measure we have of what safety actually costs at production scale. Twenty percent is high enough to be a genuine competitive disadvantage against a lab that spends nothing, which is precisely the dynamic regulation exists to correct. It also gives enterprises a useful benchmark. If your own AI deployment spends nothing on monitoring, you are running a configuration the model's own builder considers unsafe.
10. Microsoft Patches the Zero-Click CoSnitch Copilot Flaw
Microsoft patched CVE-2026-24301, nicknamed CoSnitch, on August 18. The flaw let a single malicious link auto-execute prompts and exfiltrate a user's Gmail, Drive, and Calendar data with no user interaction beyond the click. The issue was first reported in December 2024, and the fix took roughly eight months from the point it was escalated.
Zero-click means the victim does not have to approve anything, type anything, or notice anything. In an assistant with connected accounts, a prompt injected through a link inherits every permission the user granted, so the model becomes the exfiltration path rather than the target. This is the same class of problem as the GitHub Copilot Autofix incident that leaked a Snowflake Jira token, and the same as the Ray framework vulnerability CISA flagged this week. The pattern is consistent: AI tooling is being granted broad credentials before the security model around it exists.
My take: the eight-month remediation window is the part that should worry security teams more than the bug itself. Prompt injection has been a known class for three years and there is still no standard mitigation, only case-by-case patches. If you run connected AI assistants across an organisation, scope the connectors down to what each genuinely needs, and treat any assistant with mail and drive access as a data exfiltration surface, because that is what it is.
11. Cursor Ships Subscriptions, Subagents, and Long-Lived Goals
Cursor shipped a cloud agents update on August 19 adding a subscriptions system that monitors pull requests, Slack threads, and scheduled tasks, custom modes pinned in chat, subagents running on isolated VMs, a /goal command for long-lived objectives such as fixing flaky tests, and non-interrupting steering messages that queue until the agent's next tool call.
The /goal command and the subscriptions system are the same idea from two directions, which is moving agents from request-response to standing assignment. Instead of asking for a change and waiting, you give the agent an objective and a trigger, and it works when the trigger fires. Non-interrupting steering solves the practical annoyance that has made long-running agents hard to supervise, since previously correcting an agent mid-task meant stopping it. Isolated VMs for subagents address the obvious safety concern with parallel agents touching the same working tree.
My take: this is the shape agentic coding is settling into, and it matches the Linear data from yesterday showing agent-using teams at 65 weekly pull requests against 10 for everyone else. The unresolved bottleneck is still review. Standing agents that open pull requests around the clock are only a gain if someone can evaluate the output, and no vendor has solved that half. Our AI coding tools hub and AI agent frameworks hub track this space.
12. Xiaomi's Humanoid Hits 98 Percent Precision on Assembly
Xiaomi debuted a 1.7 metre humanoid robot at the World Robot Conference in Beijing on August 19, following testing at its own automotive plant. Nut-tightening precision improved from 90.2 percent to 98 percent, and folding centre-console covers reached around 90 percent. Xiaomi is prioritising smart manufacturing production lines before any consumer integration into its people, cars, and home ecosystem.
Testing a humanoid inside your own car factory is the shortest path to useful data, because you control the environment, you own the failure cost, and you already know the task takes. The move from 90.2 to 98 percent on nut-tightening is the kind of increment that decides whether a robot is deployable, since a 10 percent failure rate on a production line is unusable and a 2 percent rate can be caught downstream. The 90 percent figure on cover folding shows how much harder deformable materials remain.
My take: pair this with the Unitree listing and the picture of Chinese robotics is coherent. One company is monetising in public markets, another is quietly proving unit economics inside its own factory, and both have domestic supply chains that keep hardware costs low. Factory-first deployment is also the honest strategy. Humanoids in homes remain a demo, humanoids doing repetitive assembly are a business, and Xiaomi is saying so plainly.
13. Google Gives US Students a Free Year of Gemini Pro
Google is offering eligible US college students 12 months of Google AI Pro free, a $19.99 per month bundle including Gemini Spark, 5TB of storage, 4x higher usage limits, and Google Health Premium. More than 140 international markets get a one-year AI Plus tier with Gemini Omni and 400GB. The offer includes a Student Hub, study notebooks with diagnostic quizzes, interactive 3D visualisations, and Deep Research in Gemini Live. The redemption deadline is December 31, 2026.
Student giveaways are habit-acquisition plays and the economics are straightforward. Gemini crossed 1 billion monthly active users on August 11, so this is not about raw numbers, it is about capturing the cohort that will choose default tools for the next decade at the exact moment they are forming research and writing habits. The study-specific features matter more than the storage, because a diagnostic quiz generator and a notebook tied to course material are sticky in a way a generic chatbot is not.
My take: a free year that expires in December 2026 for redemption gives Google a full academic cycle of usage data on how students actually study with AI, which is worth more than the forgone subscription revenue. For educators the harder question arrives next term, since a tool that nudges through problems and one that hands over answers look identical from the outside. If you are teaching or training with these tools, design the assessment around the assumption that every student has them.
14. Amazon Makes Alexa+ Free and Expands Drone Delivery
Amazon auto-upgraded Fire TV devices to Alexa+, eliminating the previous $19.99 per month fee for non-Prime members, across Fire TV Sticks, Fire TV Cubes, Amazon Ember TVs, and select Hisense and Panasonic sets. New capabilities include conversational content discovery, Ring camera feeds on screen, and AI theme and rating recommendations. Amazon says Alexa+ users hold twice as many conversations as old Alexa users. Separately, Prime Air drone delivery is expanding to Chicago, Atlanta, Syracuse, Cleveland, and Boise by year end, targeting roughly 500 towns and cities, with packages up to 5 pounds delivered within 60 minutes at $3 for Prime members and $5 otherwise.
Dropping a $19.99 monthly fee to zero is a distribution decision, not a product one. Amazon has tens of millions of Fire TV devices in living rooms and the value of an assistant that people actually talk to exceeds the subscription revenue it was collecting from a small non-Prime minority. The 2x conversation metric is the number Amazon cares about, because conversation volume is what makes the assistant useful for commerce intent. Prime Air scaling from 10 metros toward 500 towns is a roughly sixfold expansion and the most concrete physical AI deployment of the week outside China.
My take: Amazon is doing to voice assistants what Google is doing to students, buying habit at the moment the underlying technology finally got good enough to keep. The Ring camera integration is the detail worth watching, because an assistant that sees your doorstep and controls your screen collects a category of household data that no chatbot does. Whether people notice that trade is a separate question.
15. Frontier Models Score 3 to 15 Percent on Hypothesis Generation
A new benchmark called Reconstruction, published in August 2026, asks models to recover a research paper's central idea from its bibliography alone, with the full text, author information, and post-publication signals stripped out. Frontier models score between 3 and 15 percent. A multi-agent pipeline using a Swiss-tournament top-four selection reached 42 percent.
The design is what makes this useful. By removing everything except the reference list, the benchmark tries to separate genuine hypothesis generation from retrieval of memorised training data, which is the confound that has made every previous claim about AI scientific creativity hard to evaluate. Frontier models scoring 3 to 15 percent suggests they are much weaker at proposing novel directions than their performance on established problems implies. The 42 percent from a multi-agent pipeline is the more hopeful number, and it says orchestration recovers a lot of what a single pass misses.
My take: this deserves more attention than it will get, because it is a rare benchmark designed to fail the models rather than flatter them. It also sits in useful tension with the Claude protein binder result from yesterday, where a model beat the industry hit rate on a real wet-lab task. The reconciliation is probably that models are strong at searching a well-defined design space and weak at deciding which space is worth searching. That distinction should shape how you deploy them in research settings.
16. European AI Data Centers Move 175km Away From the Hubs
New European AI data center projects now sit an average of 175km from major hubs, against 46km for projects built between 2022 and 2025, according to JLL data. Greenfield builds have risen to 39 percent of the pipeline from 8 percent historically. Powered land in Amsterdam, London, and Frankfurt costs about 2.36 million euros per megawatt, against 512,000 euros per megawatt in tertiary cities. Destinations include rural Spain and northern Sweden. Nvidia has separately been introducing GPU allocation holders to Nordic data center operators.
A 4.6x cost difference per megawatt of powered land explains the move on its own, but the deeper driver is that power availability, not proximity, now determines where compute can physically exist. Latency-sensitive workloads still need to be near users. Training runs and batch inference do not, so they go where the electricity is. This is the same constraint Microsoft described as a shortage of warm shells, and Pennsylvania's new permitting order is an example of the local friction pushing projects further out.
My take: the phrase worth remembering is that data centers are being brought to where power is, rather than the other way around. That inverts twenty years of colocation logic and it has real consequences for European regions that assumed proximity to financial centres was their advantage. Nvidia acting as a matchmaker between allocation holders and Nordic operators is a small detail that shows how hands-on chip vendors have become in placing their own supply.
17. The FDA and FTC Both Move on AI Oversight
The FDA published a discussion paper on August 18 proposing a two-phase evaluation framework for generative AI medical devices: nonclinical benchmarking of clinical knowledge, analytical ability, safety, and communication, followed by real-world clinical confirmation. It covers agentic systems that plan and execute multistep tasks, and it evaluates the shipped device rather than the underlying foundation model. Comments close October 19. Separately, the FTC proposed an enforcement policy statement requiring retailers who use personal data such as browsing history, location, and cart behaviour to set individualised prices to disclose that clearly and conspicuously, with a 30-day comment period open.
The FDA move lands alongside a PLOS Digital Health study finding that of 1,357 FDA-authorised AI medical devices, only 34 are linked to clinical trials and just 3 were evaluated on patient-centred outcomes such as mortality, stroke, hospitalisation, or quality of life. The study also notes that evidence comes primarily from well-resourced settings and systematically excludes pregnant women, adults over 75, and non-English speakers. Clearance signals market equivalence, not clinical benefit, and the proposed framework is a direct response to that gap.
My take: three devices out of 1,357 tested on whether patients actually did better is the statistic of the week and it should end the habit of citing FDA clearance as evidence of clinical value. Testing a generative device the way you would examine a clinician is a sensible idea and an operationally hard one, since the shipped product changes every time the underlying model updates. On the FTC side, personalised pricing disclosure is overdue, and the enforcement question will be whether clearly and conspicuously survives contact with checkout page design.
18. Meta's Trial, Smart Glasses, and Surveillance in Brief
Several stories from the same cycle are worth logging together.
● Arturo Bejar, a former Meta safety engineer with eight years at the company, testified on August 19 in the 29-state attorneys general trial that Meta's culture prioritised user counts over safety and that only Mark Zuckerberg could have changed it. He called Zuckerberg's 2021 claim that profit was not prioritised over safety inaccurate, and pointed to internal studies documenting harm to children that were repeatedly set aside.
● Teen boys in US schools have been covertly filming girls in hallways, cafeterias, and classrooms using Meta Ray-Ban AI glasses, posting the footage to TikTok and Instagram, with one account passing 64,000 followers. Districts are banning smart glasses. Meta points to the on-record LED, which can be defeated with a sticker.
● Flock Safety is testing OS Investigate, formerly Nightshift, which lets officers search for people and vehicles by movement pattern alone with no plate, name, or alleged crime required. It ships with 69 prewritten prompts and cross-references plates against case files, 911 logs, ballistics, and commercial identity records, returning names, addresses, relatives, and phone numbers.
● Munich Re is acquiring cyber insurer At-Bay for $575 million enterprise value, folding it into subsidiary HSB, with closing targeted for the first quarter of 2027. At-Bay wrote $278 million in gross premiums in the first half of 2026 and was valued at $1.35 billion in 2021, making this a visible valuation reset for cyber insurtech.
● Inco published DFlash 2, a speculative decoding update adding a path selector for coherent token sequences and a local convolution module that fixes accuracy decay at block ends. It claims 20 percent more output per verification pass at 1 percent added cycle latency, and 2.7x to 3.4x throughput against autoregressive decoding, with output quality unchanged.
● OpenAI retires o3 from ChatGPT on August 26, 2026 after a 90-day sunset, while GPT-5.4 mini rolls out to Free and Go users through the Thinking feature and as a rate-limit fallback for everyone else.
Two threads run through that list. Accountability for AI-adjacent harm is arriving through courtrooms and school district policies rather than through federal rules, and the surveillance capability being sold to police departments is expanding faster than the oversight framework around it. Both are governance stories that will outlast this week's model releases.
19. Where the Frontier Models Stand Today
Here is the practical state of the frontier as of August 20, 2026, for teams choosing what to build on.
The scheduling note for this week is the o3 retirement on August 26. If anything in your stack still pins o3, migrate before then rather than discovering it through a failed request. Full comparisons live in our best AI models ranking and the GPT-5.6 review.
20. What to Watch Next in AI
Four things from today carry into next week.
● GLM-5.3 open weights, targeted for around August 28. A model leading a cybersecurity benchmark being released openly is a genuine test of how labs handle dual-use capability.
● Unitree's second week of trading. First-day pops on thin floats correct often. Where the price settles by month end tells you the real valuation, not the 629 percent headline.
● The o3 retirement on August 26 and any migration breakage that follows.
● FDA comments on the generative device framework, open until October 19. Whoever files shapes how AI clinical tools are evaluated for the next decade.
The through-line for August 20 is that the money moved down the stack. A robot maker, a chip designer, a data-labeling firm, an inference silicon startup, and a foundry all repriced in a single cycle, while the model releases were incremental by comparison. That is what a maturing industry looks like: the value stops concentrating in the model and starts spreading across everything required to build, power, and supply it.
Frequently Asked Questions
How much did Unitree Robotics stock rise on its debut?
Unitree Robotics opened at 1,100 yuan against an IPO price of 150.8 yuan on August 19, 2026, a 629 percent gain, and closed its first session at 845 yuan, up 460 percent. The company raised 6.1 billion yuan, about $904 million, on Shanghai's STAR Market. Retail demand oversubscribed the offering roughly 5,500 times.
What is Unitree Robotics worth?
At its intraday peak Unitree reached about 445 billion yuan, roughly $66 billion. It closed its first session valued near $50 billion. It is the first humanoid robot maker listed in mainland China and shipped more than 5,000 humanoid units in 2025.
Why did Marvell stock jump?
Marvell rose more than 10 percent after granting Google a warrant to buy up to 58.97 million shares at $206.58 each, worth as much as $12.2 billion, tied to a custom chip agreement covering AI inference accelerators, storage and network controllers, and near-memory compute. Most shares vest as Google hits cumulative purchase thresholds of $500 million.
When is the OpenAI IPO?
OpenAI CFO Sarah Friar told an all-hands meeting the company will be public in 2027, or sooner if the business continues to inflect. Earlier reporting described a target valuation near $1 trillion. Anthropic has already filed confidentially and could list this autumn.
Is GLM-5.3 better than Claude or GPT-5.6?
GLM-5.3 leads CyberGym at 84.5 percent, ahead of Claude Mythos 5 and GPT-5.6 Sol, and lifts Terminal-Bench 3.0 from 4.6 percent to 28.3 percent. It is not a blanket leader. Fable 5 and GPT-5.6 Sol still score higher on raw terminal coding and general reasoning with tools. Open weights are targeted for around August 28, 2026.
What is the CoSnitch Copilot vulnerability?
CoSnitch, tracked as CVE-2026-24301, was a zero-click Microsoft Copilot flaw patched on August 18, 2026. A single malicious link could auto-execute prompts and exfiltrate a user's Gmail, Drive, and Calendar data with no further interaction. It was first reported in December 2024.
Is Google giving students Gemini Pro for free?
Yes. Eligible US college students get 12 months of Google AI Pro free, a $19.99 per month bundle including Gemini Spark, 5TB of storage, 4x higher usage limits, and Google Health Premium. More than 140 international markets get a one-year AI Plus tier. Redemption closes December 31, 2026.
Can AI models generate original research ideas?
Weakly, based on the Reconstruction benchmark published in August 2026. Asked to recover a paper's central idea from its bibliography alone, frontier models scored between 3 and 15 percent. A multi-agent pipeline using Swiss-tournament selection reached 42 percent, suggesting orchestration recovers much of what a single pass misses.
Recommended Blogs
● Anthropic Raises Its Own AI Risk Level: AI News August 19 2026
● Stripe Buys OpenRouter for $7 Billion: AI News August 18 2026
● Inside OpenAI's $1 Trillion IPO: AI News August 17 2026
● Anthropic Turns Its First Profit: AI News August 16 2026
● Best AI Models July 2026: Ranked by Use Case and Price
● GPT-5.6 Review: Sol, Terra, Luna Benchmarks and Pricing
● Kimi K3 Review: Benchmarks, Pricing, and K2 Comparison
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you're a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website: buildfastwithai.com
● LinkedIn: Build Fast with AI
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
● AI Workshops: Free resources, upcoming events, and past recordings
● Unrot: Learn AI in 5 minutes a day (free micro-learning app)
● Gen AI Experiments: free cookbooks and notebooks on GitHub
GLM-5.3's open weights and Unitree's second week of trading both land in the next few days. Follow Build Fast with AI so each recap reaches you before your standup.
References
● Unitree Shanghai debut and valuation (Fortune)
● Unitree 629 percent open (Benzinga)
● Marvell warrant to Google (CNBC)
● OpenAI CFO on a 2027 listing (CNBC)
● Nvidia weighs Mercor stake (The Information)
● Fractile raise and Anthropic order (The Next Web)
● Samsung foundry price increase (Reuters)
● GLM-5.3 benchmarks and pricing (DataNorth)
● Copilot CoSnitch vulnerability (CSO Online)
● Cursor cloud agents changelog (Cursor)
● Xiaomi humanoid at World Robot Conference (TechNode)
● Google AI Pro for students (Google Blog)
● Alexa+ free on Fire TV (TechCrunch)
● European data center relocation data (Reuters)
● FDA generative AI device framework (US FDA)
