Global Tech Weekly: The Week the AI Bills Came Due, From Memory Chips to Power Grids

Illustration: the week's stories shared one thread, which was who pays for the machines behind AI.

Some weeks in tech are about shiny new gadgets. This one was about the bills. Who pays for the chips, the power plants and the data centres, and who gets hurt when the bill lands? Grab a coffee, because there is a lot to unpack.

The chip money is getting almost silly

Start with a number that made me read it twice. Micron, the memory-chip maker, reported record fourth-quarter revenue of $54.23 billion, and $133.19 billion for its full fiscal 2026, according to its SEC filing. Memory used to be the dull, cyclical corner of the chip business. Now every AI server needs piles of it, and the sums look like a different industry.

Illustration: memory and compute chips are where the AI boom turns into revenue.

The financing is getting just as creative. A filing reported by Reuters via The Star says Broadcom will lend Anthropic up to $42 billion to lease the TPU chips Broadcom helps design. In plain terms, the chip supplier is now also the bank. That tells you how hard the industry is leaning on borrowed money to keep building, and it is worth remembering when people talk about "demand" as if it were all customers paying cash.

Then there is the geopolitics of renting. The Financial Times reported that Tencent has leased roughly 100,000 advanced AI chips from Oracle in a deal valued at around $7 billion, with the hardware sitting in Oracle's data centres in Southeast Asia, as summarised by The Next Web. A Chinese company gets access to top-end compute without owning it or shipping it home. Expect regulators in Washington and elsewhere to study that arrangement closely.

AMD's Lisa Su, meanwhile, said the company is lining up much more AI chip supply for 2027, with talks under way with TSMC, Samsung and SK Hynix, according to a daily roundup from Tech Startups. Planning three to five years ahead for factory capacity is the new normal.

Power is the real bottleneck

If chips are the glamorous part, electricity is the part that actually limits things. The same roundup reports that Google has signed a 3.59-gigawatt power procurement deal with Constellation Energy, including 890 MW from nuclear upgrades across Illinois, Pennsylvania and New Jersey, over 20 years. I could only confirm this through that one roundup, so treat the details as reported rather than settled.

Illustration: the grid, not the chip, is increasingly what decides where AI gets built.

Look at the other places money is flowing. According to the same coverage, Saudi Arabia's HUMAIN is expanding its data-centre programme from 50 MW to 250 MW. Singapore-based DayOne has filed for a Nasdaq listing, and the Tech Startups write-up says one short-video customer makes up about 69% of its revenue. In Japan, a partnership of JERA, Dell and others is planning a 400 MW AI site at a thermal power station. Taiwan's Foxconn is reported to have passed its first NT$1 trillion sales month on AI servers. Different continents, same story: the data-centre build-out is now an energy and real-estate business as much as a software one.

One cautionary detail from that DayOne filing: when one customer is nearly 70% of your revenue, your fortunes are tied to theirs. Worth keeping in mind for anyone whose own employer is betting heavily on a single AI vendor.

Rules arrive for the workplace

Not everything this week was about building. In California, Governor Gavin Newsom signed a set of laws aimed at AI in the workplace, according to CTV News and the CP24 wire report. Employers can't use AI with biometric data to predict a worker's emotional state. They must give written notice when AI is behind mass layoffs. And they can't rely on AI alone to decide whether to fire someone.

Illustration: new rules are starting to put a human back in the loop.

Elsewhere, privacy rules are hardening around wearables. Norway is reportedly proposing temporary limits on AI smart glasses in certain public spaces over covert-recording worries, and Italy's prime minister reportedly filed to protect her own voice as an EU trademark. Both come from the 5 October roundup, so check the primary announcements before you quote them.

There was a sobering security note too. The roundups report a breach at Denmark's central person register exposing 8.8 million records, and a wave of attacks on seven South Korean financial institutions affecting more than 67,000 people. If you work anywhere near customer data, this is your weekly reminder to check who has access to what.

Gadgets, and a free tier shrinking

On the consumer side, the news is mostly rumour and anticipation. Apple is widely reported, via 9to5Mac, to be planning a smart-home hub, a HomePod mini and a new Apple TV around 13 October, with Macs and a possible foldable iPhone to follow later in the month. None of that is confirmed by Apple, and the report itself rests on unnamed sources, so file it under "likely, not certain".

Illustration: October is shaping up as a busy month for devices, if the rumours hold.

The more concrete change for everyday users is on the software side. The 6 October roundup says Google will, from 9 October, limit free Gemini users to its Flash Lite model, with the standard Flash model moving to a $4.99-a-month plan. I couldn't find a second source in the time I had, so check Google's own announcement before you change your plans. Still, it fits the pattern: as the costs above pile up, free tiers get thinner and ads and subscriptions fill the gap. OpenAI, for instance, is reported to be testing labelled visual ads inside ChatGPT's image generation for some US users.

What it means for you

Two takeaways for working people.

First, ask what happens if your tools get more expensive. The industry is financing its growth with debt, leases and ever-bigger power contracts. That works while demand keeps climbing. If you're building a workflow, a team or a product on top of a free or cheap AI tier, assume the price may change and keep a Plan B.

Second, expect more rules about people, not just models. California's laws are a template other places may copy. Whether you manage a team or sit on one, find out now how your employer uses AI in hiring, performance reviews and layoffs, and whether a human is truly accountable for the decision. That conversation is easier to have before a regulator asks for it.

The short version

Chips are making fortunes, power is the new constraint, and the bills are being passed around through loans, leases and subscriptions. Meanwhile, lawmakers have started writing rules for how AI treats workers. Nothing here is a crisis, but all of it is worth watching.

A note on sourcing: figures for Micron, Broadcom/Anthropic, Tencent/Oracle and the California laws were checked against the sources linked above. Items drawn only from daily news roundups are labelled as reported. The illustrations are original artwork made for this edition.