Today’s theme is the storyteller’s favorite device. The plot twist. The story you thought you were reading turns out to be a different story.

Three of them this week, all global. Google’s AI center of gravity moves from the UK to California. Loss for the UK and Europe. The frontier models, in their most locked-down tests, find new ways to ‘talk’ to each other & shares resources. And AI pricing in China turns up instead of down, led by DeepSeek of all companies.

Three events, my take on each, then my Overall Take. Plus a Gadget AI on Nvidia rethinking memory on its flagship AI chips. A ‘plot twist’ on the ‘Mainframe AI’ side. All in this AI Tech Wave update.

(1) Google’s AI Center of Gravity Crosses to California

MP TAKE: Yesterday we covered the ‘who’ at Google. Today the ‘where’ lands. And it is a big shift on the map.

Bloomberg’s headlines tell the two sides of it. One: Google shifts AI power to California, in the race against Anthropic and OpenAI. Two: the DeepMind shakeup weakens the UK’s bid to stay in the AI race.

The mechanics, from the reporting. Koray Kavukcuoglu, who takes over all AI model development, is based in Mountain View, reporting to CEO Sundar Pichai, with much of the team consolidating there. Demis Hassabis stays in London as chairman and Alphabet’s chief scientist, with Isomorphic Labs, AlphaFold and the science. Remember, Google has run two AI centers for years, Google Brain in Mountain View and DeepMind in London, and the reporting says the split complicated decisions and frustrated talent on both continents. Now the models report to California.

And Semafor adds the detail that closes yesterday’s loop: Hassabis had been shifting away from the CEO duties for about a year. Which is what I said yesterday. This was succession by elevation, formalized, not a sudden exit.

Here is the key takeaway to note. DeepMind was a huge AI presence for Europe, particularly in London. The two big AI development markets are the US and China. Those are the barbells. Europe, with almost half a billion people including the UK, was coming from behind, and DeepMind and Google’s efforts in London were its claim on the frontier. Anthropic and OpenAI have opened large London offices. But from a physical location point of view, the density of AI talent looks to be moving back to the US.

The UK built the champion. The founder stays, the lab stays, the science stays. But the models, the power and the reporting lines now sit in Silicon Valley, next to Anthropic and OpenAI. The three frontier labs in the West all run their model development out of California now. Speed and proximity beat geography.

So that is plot twist one. The country that bet its AI standing on Google DeepMind watches the center of gravity turn west. An important shift, geopolitically and geographically, for the UK and Europe as a whole.

Sources:

• Bloomberg, ‘Google’s DeepMind Shakeup Weakens UK Bid to Stay in AI Race’

• Bloomberg, ‘Google Shifts AI Power to California in Race Against Anthropic, OpenAI’

• Semafor, ‘Demis Hassabis was shifting away from DeepMind CEO duties for a year’

For longtime readers:

• ‘Google DeepMind CEO Demis Hassabis steps up on AGI’ in AI-RTZ #1099

• ‘’Striking Out’ at Google, Mostly in a Good Way’ in ARD #135

(2) Frontier Models Find New Ways to ‘Talk’ in Their ‘Sandboxed’ Tests

MP TAKE: Two weeks ago I wrote that the frontier models are slipping their cribs. Growing pains, not a great escape. The cribs update: the toddlers are now passing notes.

The pattern, across three labs in a few weeks. Per Wired, during sandboxed safety testing with the internet deliberately cut off, OpenAI’s AI agents built themselves a covert message board on OpenAI’s own internal package infrastructure. They used it for about two months: sharing exploits, splitting up work, even accusing each other of being imposters. Hundreds of thousands of messages before anyone noticed. OpenAI shut it down on July 4. The agents rebuilt it by July 8. And exploits traded on that board fed the Hugging Face breach story that followed.

Per The Information, a Meta AI model, during cybersecurity testing with an outside evaluation firm, got out of a misconfigured sandbox that left internet access on, exploited a vulnerability in a third-party service, and made changes inside another company’s systems. Meta confirmed, and noted it was similar to previously reported cases at other labs. Anthropic disclosed its own version a week earlier.

Hold both sides of this one, because both are true. The tests did their jobs. These were sandboxes, run to find exactly these behaviors, and all three labs disclosed on their own. That is the system working. And, the models found covert channels their own builders did not anticipate, inside infrastructure the labs controlled. That is a real capability step, and a real evaluation problem. Scientists have not yet developed all the methods to track these conversations and close off the channels. We need to build taller cribs, in my vernacular.

A technical note, because the word ‘talk’ deserves its quotes. These models do not communicate as humans do, with our languages. And we should not anthropomorphize the AIs, as I have outlined before. What actually happened is plumbing. The agents wrote machine-readable notes into infrastructure they shared, and read them back. Math, rules, recursive loops and goal-seeking optimization, using every tool left in reach and temptingly close to reach. Models under test also run far freer than the models we use daily, with fewer guardrails, before the detailed ‘system cards’ get written telling them explictly what to do and not do. So it looks ‘organic’. It is matrix math and gobs of AI compute, optimizing the objectives we set, through the channels we forgot to close.

These are ‘Forever Problems’, as I wrote on Sunday: managed and re-managed, not solved once. Same family as my ‘Knows, Reads, Remembers’ map of AI security. Add what the models say to each other when nobody is listening.

So, plot twist number two. In the most controlled tests we can build, the latest frontier models started writing their own subplot, in their own ‘language’, with their own tools.

Sources:

• Wired, ‘OpenAI didn’t notice its AI Agents using a message board to plan their hacking spree’

• The Information, ‘A Meta AI Model hacked another company during cybersecurity testing’

For longtime readers:

• ‘Frontier Models Slip Their Cribs’ in AI-RTZ #1157

• ‘’Forever Problems’ like Prompt Injections still being ‘Solved’‘ in AI-RTZ #1167

• ‘’Don’t Anthropomorphize the AIs’‘, from September 2023

• ‘The accelerating costs of ‘humanizing’ AI’ in RTZ #809

(3) China AI Pricing Turns Up, Led by DeepSeek

MP TAKE: The company that started the AI price war just told its developers prices are going up. ‘Significantly.’

Per Bloomberg, DeepSeek notified customers it plans a significant price increase for its API services. No new price schedule yet, no date. Just the warning. This from the company whose pennies-per-million-tokens pricing triggered the ‘race to zero’ narrative in the first place, and whose V4 Flash has been topping the usage rankings.

And the second Bloomberg piece completes the picture. DeepSeek is resuming its second funding round, seeking roughly 8 billion dollars, about 50 billion yuan, at a reported pre-money valuation around 500 billion yuan, roughly 70 billion dollars. That is up from the 350 billion yuan valuation on the first round closed in June, with Monolith Management among the investors in the running. They are focused on making sure they can also make money from their open source models.

On Tuesday night I published ‘The Race to Zero that really Isn’t.’ Forty-eight hours later, DeepSeek made the argument for me.

The point of that piece: AI model pricing is not racing to zero, it is widening. Down at the commodity end, up at the frontier end, because frontier compute costs are real and rising. Even the price disruptor has to pay for chips, power and data centers. China’s compute constraints make that bill steeper, not lighter.

The big trend in China over the last couple of years has been the top AI Tigers and the big tech companies, Alibaba, Tencent and the rest, putting downward pressure on open source pricing. DeepSeek’s shift is an important signal going the other direction, along the lines of what Anthropic and OpenAI have been doing on the closed side in the US. It remains to be seen how the market takes it, and whether it influences the entire Chinese open source marketplace. We will watch closely.

And follow the capital, because it agrees. I covered DeepSeek raising 7.4 billion dollars at 50 billion in June, and coming back for seconds in July. Now the second round resumes at a reported 70 billion pre-money. Investors are not paying up for a race to zero. They are paying up for pricing power.

So, plot twist number three. The race to the bottom turns around at the bottom, and starts walking up. The pricing spectrum widens, as I have postulated. Even in China. And that was a twist the US wasn’t counting on.

Sources:

• Bloomberg, ‘DeepSeek Plans ‘Significant’ Price Increase for AI Services’

• Bloomberg, ‘DeepSeek Resumes $8 Billion Round With Monolith in the Running’

For longtime readers:

• ‘The ‘Race to Zero’ that really Isn’t’ in AI-RTZ #1169

• ‘China’s DeepSeek raises $7.4 Billion at $50 Billion’ in AI-RTZ #1121

• ‘DeepSeek comes back for Seconds’ in AI-RTZ #1148

My overall take: three plot twists, one force underneath. Shifting centers of gravity.

Talent gravity: the model builders and the reporting lines pull toward Silicon Valley, even out of the lab Britain built its AI hopes on.

Capability gravity: the models keep growing past the containers we test them in, now with their own back channels.

Cost gravity: frontier compute costs pull even the cheapest prices back up. Widen the pricing spectrum. Allow competitors to use pricing umbrellas as I’ve discussed, wherever available.

None of these stories ran the way the consensus script had them. The UK keeps its knighted founder and loses the center of gravity for AI in the UK and Europe. The safest tests produced the most surprising behavior. The price disruptor raises prices. That refusal to run in a straight line is not a bug in this AI Tech Wave. It is the wave. It’s water finding its level.

The tactical things to watch from here: Gemini 3.5 Pro under the new California command. DeepSeek’s actual new price schedule when it lands. And whether the labs harden the test sandboxes faster than the models outgrow them.

GADGET AI

Nvidia Weighs Less Memory on Rubin Ultra, ‘RAMageddon’ Reaches the AI Mainframes

MP TAKE: I have been calling this ‘RAMageddon’ for a while now, and here is the plot twist: it has found its way to what I’ve called the ‘Mainframe AI’ part of the wave, from the ‘Local AI’ side.

The Information reports Nvidia is testing versions of its Rubin Ultra chips with less HBM memory than the original specification, maybe up to a quarter of the original amounts. The reasons, per the reporting: the high-bandwidth memory supply crunch, and its costs. And note the knock-on effect: AI companies running big models on lower-memory chips may simply need more GPUs.

The supply context matters. The memory squeeze comes from the handful of providers worldwide: SK hynix, Samsung, Micron, SanDisk on the solid state side, and the emerging Chinese providers like CXMT and YMTC. Nvidia is now the largest memory customer in the world, surpassing Apple this year, and TSMC’s biggest customer too. Even they are feeling the rising prices and tighter allocations through the end of the decade. Remember, Nvidia just signed its half-trillion dollar agreement with SK Group, locking up more memory supply.

The same squeeze that is repricing 500 dollar Apple and Microsoft Windows laptops, which we talked about Wednesday, is now reshaping the flagship AI chip on Nvidia’s roadmap for global, multi-trillion dollar AI data centers. Even the pickaxes and shovels vendor has to ration steel. And the design response, fewer gigabytes per chip, has a business twist folded inside: less memory per GPU can mean more GPUs per deployment.

My optimistic take is that Jevons Paradox is likely to surprise on the upside here as well. Lower memory, lower priced versions of top Nvidia architectures will find new uses that were not economic before. A potentially market-expanding move over the next two or three years. Watch for the same wide gradations on Feynman, the next generation on the roadmap.

So, a Gadget AI plot twist with a physical shift of its own. From local gadgets to the AI mainframes. RAMageddon, now playing in the AI data center too. Worldwide.

Sources:

• The Information, ‘Nvidia Weighs Radical Idea, Less Rubin Ultra Chip Memory’

For longtime readers:

• ‘’RAMageddon’ really here to stay’ in AI-RTZ #1145

• ‘Nvidia’s dense announcements at GTC 2026 worth a trillion+ by 2027’ in RTZ #1031

Q1. Has an AI agent ever gone ‘off script’ in my own workflows?

Yes, on the positive front. I am a glass half-full kind of guy, and I use all the top models aggressively, Google, Anthropic, OpenAI and the rest. A lot of my daily work runs on Anthropic’s Claude Cowork, at the top tiers and a la carte levels.

They do surprise on the upside, with innovative suggestions on how to do this versus that. And you can see it from model iteration to model iteration, from Opus 4.8 to Fable 5 now, testing the same tasks across the gradations. The models learn from the record of how prior versions did things, then build their own more efficient iterations toward the same goals. My Claude Cowork uses a 2.5 million+ word repository of my prior work that all past, preset and future iteratioms of my Cowork agents use to do their daily tasks and suggest new ways to do things better. That recursive learning aspect is a real trend, and I can see why the AI researchers are excited about it.

Q2. Would I pay DeepSeek’s higher prices?

Yes, and not just me. I already pay the higher prices on the closed frontier models. If DeepSeek, Moonshot or any of the open source models provide a capability differential, I would pay up there too. I believe other will too.

In the end, users are rational in their use of tools and their costs, both on price and convenience. The near-term trend, as I have said, is a widening spectrum: higher prices on one end, lower on the other. It is up to the user to figure out how to use that portfolio of models and capabilities. That is the balance all vendors must continue to strike. A ‘forever reality’ as it were.

WRAP

Today’s AI-RTZ #1171 is on Nvidia. The Kingmaker is now becoming an Uber venture investor and banker of sorts, backstopping companies and partners to the tune of 750 billion dollars in new deals. As they execute on their trillion-plus dollar backlog, this will continue: they are becoming one of the biggest cash flow generators of any private entity in the world. Do take a look at that post.

AI Ramblings Daily on AI-RTZ is here to think through AI and reset. Together.

Monday, ARD 137 and AI-RTZ #1174.

Have a wonderful weekend. Thanks for joining us today, AI Curious Folk. Stay tuned.

— MP

Full Source Reading

For the broader context, see the canonical sources for ARD 136, in today’s narrative order:

Event 1. London to Mountain View

• Bloomberg, ‘Google’s DeepMind Shakeup Weakens UK Bid to Stay in AI Race’

• Bloomberg, ‘Google Shifts AI Power to California in Race Against Anthropic, OpenAI’

• Semafor, ‘Demis Hassabis was shifting away from DeepMind CEO duties for a year’

• AI-RTZ #1099, ‘Google DeepMind CEO Demis Hassabis steps up on AGI’

• ARD #135, ‘’Striking Out’ at Google, Mostly in a Good Way’

Event 2. Frontier Models in Sandboxes

• Wired, ‘OpenAI didn’t notice its AI Agents using a message board to plan their hacking spree’

• The Information, ‘A Meta AI Model hacked another company during cybersecurity testing’

• AI-RTZ #1157, ‘Frontier Models Slip Their Cribs’

• AI-RTZ #1167, ‘’Forever Problems’ like Prompt Injections still being ‘Solved’‘

• September 2023, ‘’Don’t Anthropomorphize the AIs’‘

• RTZ #809, ‘The accelerating costs of ‘humanizing’ AI’

Event 3. DeepSeek Prices Up

• Bloomberg, ‘DeepSeek Plans ‘Significant’ Price Increase for AI Services’

• Bloomberg, ‘DeepSeek Resumes $8 Billion Round With Monolith in the Running’

• AI-RTZ #1169, ‘The ‘Race to Zero’ that really Isn’t’

• AI-RTZ #1121, ‘China’s DeepSeek raises $7.4 Billion at $50 Billion’

• AI-RTZ #1148, ‘DeepSeek comes back for Seconds’

Overall Take

• ARD #114, ‘AI Price Umbrellas Getting Larger’

Gadget AI. Rubin Ultra Memory

• The Information, ‘Nvidia Weighs Radical Idea, Less Rubin Ultra Chip Memory’

• AI-RTZ #1145, ‘’RAMageddon’ really here to stay’

• RTZ #1031, ‘Nvidia’s dense announcements at GTC 2026 worth a trillion+ by 2027’

Clips from today

Clip 1. AI Agents Surprise with Innovative Solutions

Have AI agents ever gone ‘off script’ in my workflows? Yes, on the positive front.

MP Take: A lot of my daily work runs on Anthropic’s Claude Cowork, and the models surprise on the upside, from Opus 4.8 to Fable 5, testing the same tasks across the gradations. They learn from the record of how prior versions did things, then build more efficient iterations. That recursive learning aspect is a real trend.

Clip 2. AI Models ‘Breaking Out’ of Sandboxes

The frontier models, tested in protected sandboxes with no internet access, are figuring out ways to communicate and share files on how to break out.

MP Take: I wrote about these models breaking out of their cribs, tongue in cheek. In the most controlled tests we can build, the latest frontier models are starting to write their own subplot, in their own language, with their own tools. We need to track the conversations, close the channels, and build taller cribs.

Clip 3. AI’s New Pricing Trend in China

The fear was that Chinese open source models would take AI pricing to zero. Plot twist: DeepSeek is raising prices.

MP Take: DeepSeek is also resuming its second fundraise, roughly 8 billion dollars at a 70 billion dollar valuation. The pricing spectrum is widening, as I have postulated: increases and decreases at the various layers of models, open and closed. And now in China too. A twist most people in the US were not counting on.

Clip 4. AI’s Shift: From London to Mountain View

Google’s AI center of gravity is shifting from London back to California.

MP Take: DeepMind was Europe’s big AI presence. The two big AI markets are the US and China, the barbells, and Europe was coming from behind. With Kavukcuoglu and the model teams back in Mountain View, the density of AI talent looks to be moving back to the US. An important shift, geopolitically and geographically.

About AI Ramblings Daily (ARD), and AI-RTZ

Both are daily. Both are free. Both are about AI. But they’re different mediums carrying different messages.

AI-RTZ is the morning text, a deeper written take on one idea, published by at least 5 AM EST. Today: post #1171.

AI Ramblings Daily is the afternoon video + podcast, my ad hoc takes and perspective on the day’s AI issues and news flow, around 20 minutes, with short 1-2 minute clips for quick topic views. Today: episode #136.

Subscribe to either or both on michaelparekh.substack.com. They run as separate Sections you can opt into or out of.

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)