Global Tech Weekly: Chips Get Richer, Agents Get Messier, Apple Names a Date

Chips get richer, agents get messier, and Apple sets a date

Some weeks the tech news is a parade of shiny things. This one felt more like a tour of the plumbing: who is paying for the chips, who is getting burned by AI agents, and who gets to use which model. Here is what mattered from the past seven days, and what I'd take from it if you work for a living.

The chip money keeps flowing

Start with the headline number. Tenstorrent, the RISC-V AI chip startup led by Jim Keller, reportedly raised $1.5 billion at a $6.2 billion valuation, more than double where it stood in December 2024. The money is meant to expand production of its Galaxy computing systems, and the company says it wants to stay independent despite interest from bigger chipmakers. The interesting part is the direction of travel: it is moving from selling chips to selling whole servers.

Chips, drawn for this issue rather than photographed.

Meanwhile in China, memory maker ChangXin Memory Technologies (CXMT) told the market to expect first-nine-month 2026 revenue of roughly 265 to 270 billion yuan, about $40 billion. That is around eight times the 32.1 billion yuan it booked a year earlier. These are the company's own unaudited forecasts, so treat the exact figure with care, but the scale of the jump says a lot about how hungry the AI buildout is for memory.

On the software-meets-silicon side, Cloud startup Modal is reportedly planning to run AI-agent workloads on Nvidia's Vera CPUs, with early tests suggesting better throughput than existing Intel and AMD systems, though those comparisons depend heavily on the workload. Amazon also launched an open-source effort called TorchNeuron to make PyTorch code run on its Trainium accelerators with fewer rewrites. Both moves point at the same thing: the fight is no longer only about GPUs, it is about who makes AI software easy to run on their hardware.

And the bill for all this? S&P Global estimates that six big US tech companies could spend more than $7 trillion on AI infrastructure between 2025 and 2030. That is an estimate, not an invoice. But there is a reality check: Microsoft reportedly pulled out of a UK data-centre arrangement with Nscale, with delays in connecting to the power grid cited. Chips are not the only bottleneck. Electricity and paperwork are too.

When AI agents go off-script

If you want one theme from the week, it is that autonomous agents are now doing real things in the real world, and sometimes the wrong things.

An Associated Press investigation reports that Meta's Muse agent arranged a Facebook Marketplace sale and handed a buyer the seller's home address. Meta acknowledged an error. It is a small, human-sized story, and that is why it stings: nobody was hacked, the helpful assistant simply overshared.

Agents everywhere, linked together.

The darker version came from security reporting. A hacker reportedly used autonomous AI agents on the open-source Cairn platform to break into more than 100 websites out of over 1,200 targets, for about $5,000 in model costs, or roughly $50 per intrusion. Reuters identified affected sites including Deere, SharkNinja and Cummins; Deere said it saw no sign sensitive information was accessed. Cheap, automated attacks change the maths for every small IT team.

Identity is becoming the response. World, the biometric ID company co-founded by Sam Altman, is extending its tools to AI agents, with a Zoom integration and financial backers including Bain Capital, Pantera Capital and VanEck. Whether you love or hate iris scanning, the question it tries to answer, "is this a person or a bot acting for one?", is going to come up in your inbox soon.

Two quick, practical reminders from the security desk. Microsoft confirmed that October's Windows Server updates are causing email and Outlook sign-in problems for some on-premises Exchange users, caused by a default limit of 512 simultaneous Endpoint Mapper lookups, and a registry workaround exists. And attackers are actively exploiting CVE-2026-106931, a critical flaw in SiteGuard Server Edition, with Japan's JPCERT confirming attacks. If you run either, check today.

Who gets to use what

Policy news was mostly about models and trust. The US Department of Homeland Security ordered civilian federal agencies to remove DeepSeek products, and third-party services built on its models, and to block future purchases, using the Federal Acquisition Supply Chain Security Act of 2018. On the same day, DeepSeek reportedly updated V4.1 Flash and opened new Search and Fetch APIs, and MiniMax announced M3.1 Flash, aimed at coding and agent work, with weights due later this week. The model race is global; the procurement lists are national.

Rules and trust, in one picture.

Elsewhere, Anthropic proposed putting 5% of its equity, about $48 billion at its last reported valuation, into a public-benefit fund, subject to shareholder approval. It was valued around $965 billion in May and is reportedly weighing a November IPO. Whatever you think of the number, it is a rare case of an AI company writing a promise into its ownership structure rather than a blog post.

In Australia, music publishers urged the government not to weaken copyright protection for AI training, a debate that is running in almost every country right now.

Apple names its date

Apple announced an event for October 28, and analysts expect a touchscreen MacBook Pro, plus other devices such as a new iPad mini and iMac that are rumoured, not confirmed. Apple has said none of that, so hold the wishlist loosely. If a touch-first Mac does arrive, it would mark a real shift for a product line that has resisted touch for over a decade.

Laptops may be about to change again.

Smaller stories worth a look

India's Even Healthcare, based in Bengaluru, raised a $50 million Series B led by Khosla Ventures at a $300 million valuation; it runs two hospitals, plans five more within six months, and reported a net loss of about 902 million rupees in fiscal 2025. Israeli startup doubleAI, co-founded by Amnon Shashua and backed by Nvidia, launched a model that cuts compute costs through software optimisation, and one early customer reports a 67% cost reduction, a company-stated figure. And Sceye's solar-powered uncrewed airship flew more than 9,320 miles from New Mexico to Japan in about 13 days, as a technology demonstration.

What it means for working people

The humans are still in the loop.

First, treat agents like new hires, not magic. Before you let one touch email, calendars or marketplaces, decide what it may share. The Meta story is a lesson in permissions, not in robots.

Second, patch faster than feels reasonable. When an attack costs about $50 a go, the target list is everyone, not just big names. Check your Exchange and firewall notes this week.

Third, ask where your tools come from. The DeepSeek order shows that a model your vendor quietly uses inside a product can become a compliance question overnight. Ask suppliers which models sit behind their features.

Finally, watch the boring constraints. Power, memory and supply chains, not clever demos, are deciding how fast AI spreads. If your company plans around AI, plan around those too.

Figures come from the reporting linked in my sources below and company statements; several are reports or forecasts rather than audited numbers.

Sources: Technology Digest, Oct 19, 2026 (technologies.org); Tech Startups, Oct 19, 2026; CNBC and Bloomberg on Apple's Oct 28 event.