Character.AI goes bankrupt and Nvidia's China share falls to 34%

A startup Google valued at roughly $2.7 billion two years ago just filed for bankruptcy, and its remains are priced at $20 million. Nvidia shipped 44% fewer accelerators into China than a year ago. Microsoft says its own silicon will run its AI from 2027. And the IMF says the money being spent on all of this is what is holding up global growth. Four stories, one thread: the numbers everyone quotes are rarely the numbers that decide anything.

Character.AI files for bankruptcy and Disney bids $20 million for the pieces

Character Technologies filed for Chapter 11 on Monday, October 12, in the U.S. Bankruptcy Court for the District of Delaware. Disney Worldwide Services enters as the stalking-horse bidder: $20 million for a perpetual, non-exclusive license to certain intellectual property and certain company claims, plus transition services staffed by former Character.ai employees Disney has already hired.
A court filing puts assets and liabilities between $50 million and $100 million. In October 2024, Google paid roughly $2.7 billion to license the technology and absorb the founders and much of the team; what is left is now priced at about 0.7% of that figure. The platform keeps running during the sale, which should close in late November or early December 2026 under Section 363 and stays open to higher offers. If no buyer agrees to operate it, it shuts down.

Why it mattersWhat it means for Latin America: Character.ai had tens of millions of casual users, a good share of them Spanish-speaking teenagers, and what Disney is buying is a license to intellectual property, not a promise to keep the service alive. If you built a workflow, a bot or a community on top of a consumer AI app, export your data now and assume the product can vanish. The harder lesson for anyone building here: the moat was the people, and when the people left for Google, $2.7 billion of value walked out with them.

Nvidia's China share falls to 34% while Huawei nearly doubles to 28%

IDC data reported by Reuters shows Nvidia shipped 678,000 AI accelerators in China in the first half of 2026, down 44% year over year, for a 34% share. Huawei shipped 572,000 units for 28%, up from just over 300,000 a year earlier. Alibaba's T-Head unit holds 7% and Cambricon 5%. IDC did not say whether smuggled B200 and B300 chips are counted.
The total Chinese market reached 2.008 million accelerators. GPU shipments fell 26% to 991,000 while non-GPU accelerators rose nearly 80% to 1.016 million, the first half in which Chinese buyers took more non-GPU silicon than GPUs. Huawei launched more than 10 chipsets in September, and DeepSeek has partnered with it on open-source software for Ascend chips, aimed squarely at Nvidia's CUDA lock-in. Jensen Huang said back in May that Nvidia had 'largely conceded' the market.

Why it mattersWhat it means for Latin America: nobody here buys accelerators by the thousand, but everybody here pays the price of inference, and that price is set by how much competition Nvidia has. CUDA is the real monopoly, not the hardware, and DeepSeek porting open-source tooling to Ascend is the first serious crack in it. Watch two things: whether cheaper non-Nvidia inference starts showing up in the per-token prices you pay, and whether Huawei's AI hardware begins arriving in the region the way its telecom gear did.

Microsoft calls Maia 300 its 'biggest bet yet' and refuses to rent it out

Microsoft unveiled Maia 300 on Monday, October 12: its third-generation in-house AI accelerator, built for inference on long conversations, reasoning and multi-step agents. It claims 40% more tokens per megawatt than unnamed top GPU systems and up to 7x the per-rack output of Maia 200. Each rack holds 96 chips at about 167 kW; each chip carries eight stacks of HBM4 and compute chiplets on TSMC's N2P process, with 288 GB of HBM4 and 18 TB/s of bandwidth per chip.
Microsoft plans gigawatt-scale deployments from 2027, making Maia the primary system running AI across its products, and unlike Google and Amazon it does not rent these chips to outside customers. Analysts expect Maia to cover roughly 10% of Microsoft's AI chip spending by early 2027. For comparison, Nvidia's GB300 rack holds 72 GPUs at about 142 kW. Rani Borkar, who runs Azure hardware, framed the goal to Reuters as 'more intelligence for every watt, every dollar.' More details come at Ignite next month.

Why it mattersWhat it means for Latin America: the competition among the big clouds has shifted from who has the best model to who pays the least per watt, and that is the number that eventually lands on your invoice. Two practical reads. First, if Microsoft, Google, Amazon and OpenAI all run their own silicon, Nvidia stops being the single toll booth and API prices have room to keep falling, as the recent Haiku cut already showed. Second, 167 kW per rack is a reminder that the binding constraint on AI in this region is not talent or capital, it is electricity.

The IMF says AI spending is what is holding global growth at 3%

IMF managing director Kristalina Georgieva said AI infrastructure spending is what keeps global growth near 3% for 2026 despite high energy prices. Top U.S. hyperscalers alone have committed about $835 billion, and AI hardware and related products now account for more than 10% of world goods trade. The capex is concentrated in a handful of economies, with the U.S., China and India as the key producers and most of the remaining top-10 AI-trade economies being Asian suppliers of chips, memory, chipmaking equipment and robotics.
She repeated her 'slow-moving tsunami, but tsunami none the less' warning on labor: the IMF estimates AI will augment, downgrade or eliminate about 40% of jobs worldwide and roughly 60% in advanced economies, with new graduates and young workers hit first, and it also threatens government tax revenue. Separately, the IMF's October Global Financial Stability Report names an AI-driven market correction among three near-term risks, estimating it could cut global GDP by up to 0.4% by 2028. Done well, the IMF says AI adoption could add up to 0.5 points of annual global growth over time.

Why it mattersWhat it means for Latin America: this is the most uncomfortable number of the day for the region. The growth comes from building the hardware, and Latin America does not build it, so it gets the labor disruption without the capex boom. The 40% jobs figure lands hardest exactly where the region has grown fastest: outsourced services, support, back office, junior development, translation, design. The hedge is not to avoid AI but to sell the output instead of the hours, and to watch whether the data-center buildout in Mexico, Brazil, Chile and Queretaro turns into local jobs or just local electricity bills.

Russian bot traffic collapsed 99.6% after drone strikes, and almost everyone read it wrong

Cloudflare data shows automated HTTP requests from Yandex's network (AS13238) fell to 0.4% of their two-week peak after Ukrainian drone strikes on Russian data centers. Yandex confirms damage at three sites: Sasovo and Vladimir fully shut down, Kaluga with equipment left unusable. Zelensky referenced the Sasovo strike on October 8 and the Kaluga campus went offline on October 11, taking dozens of Yandex Cloud services with it.
The viral reading, that Russia's bot army had collapsed, does not hold. The chart measures automated web crawling from one network, mostly search indexing, not social media posting, and a bot score estimates whether traffic is automated, not whether it is propaganda or who runs it. Cloudflare's Russia-wide totals barely moved for the week. Influence operations typically run on residential proxies, compromised IoT devices, commercial VPNs and Western cloud providers, because platforms have blocked known Russian corporate data centers since 2022. Yandex traffic started recovering on Sunday.

Why it mattersWhat it means for Latin America: this is a free lesson in how a real chart becomes a fake conclusion, and it is the same mechanism behind most of what arrives in a WhatsApp family group. The data was genuine, the source was credible, and the interpretation was still wrong, because nobody asked what the number actually counted. Before you forward the next graph, ask three questions: what exactly is being measured, over what window, and does it support the claim in the caption. Most of the time the answer to the third is no.

The big picture

Four of today's five stories are about the gap between a headline number and the number that actually decides. Google's $2.7 billion valued Character.ai's people, not its product, and the product is now worth $20 million without them. Nvidia's share in China fell not because its chips got worse but because the rules changed and a competitor got good enough. Microsoft's pitch for Maia 300 is not a benchmark score, it is tokens per megawatt, because watts are the binding constraint now. The IMF's 3% growth is real and so is its 40% jobs figure, and they describe the same spending from opposite ends. And a perfectly accurate Cloudflare chart produced a conclusion that was flatly false. The useful habit this week is not finding better numbers, it is asking what each one is counting.

Go deeper

Sources & further reading

Topics:AI Chips Big Tech Startups