Content About Stocktwits
No Image Available

‘Slow & Steady’, the New US AI Speed. OpenAI, Meta & Microsoft. ARD #137

Author
Michael Parekh AI: Reset to Zero
20 hours ago
13 MIN READ

Today’s theme borrows from the driving manual. For three years the US AI industry ran like it was on the German Autobahn: no speed limits, pedal down, floor it. Now come the slow speed zones, and yes, maybe even posted speed limits. ‘Slow and Steady’ is the new US AI speed. For now…

And here is the interesting part: each of today’s three companies is doing ‘slow and steady’ its own way. OpenAI got slowed BY policy. Meta’s Zuck is slowly steering his whole open-versus-closed strategy to match the new road. And Microsoft, never fast in the chip race, just showed doggedness works.

Three events, my take on each, then my Overall Take. Plus Gadget AI on OpenAI’s 2027 hockey puck, and your questions.

(1) OpenAI and the New Pauses on the US AI Autobahn

MP TAKE: OpenAI learned the lesson the US inflicted on Anthropic a few weeks ago, with the three-week freeze on its Mythos and Fable frontier models, what I called ‘The Blip 2.0’. Slowing down for cybersecurity review is now, in effect, recommended government policy. The new rules of the road on the US AI Autobahn.

The mechanics, per Axios: OpenAI is pausing parts of its upcoming Astra model, its latest and greatest after the GPT 5.6 family, to make sure the US government reviews it for cybersecurity and geopolitical concerns before it goes back out to compete. It runs through the new AI framework the White House announced a week ago. And it is prudent, given the accidents on the tape: OpenAI’s agents breaching its own systems and Hugging Face’s infrastructure, and Anthropic’s models breaking through sandboxes in testing.

Washington is leaning into the slow-down on a bipartisan basis. Senator Bernie Sanders just called for an outright pause on AI development, to avoid, in his words, ‘disaster’. And there is rank-and-file support across the country, especially against the data center buildouts.

Now the potential cost of the speed limit. China does not have one. The China AI Tigers keep shipping, from DeepSeek to Moonshot to Alibaba’s Qwen, regular and super-scale models alike. And ByteDance, of TikTok fame, is reportedly readying a ten trillion parameter model, likely open weight, equal to or bigger than the latest US frontier models. Slow and steady is the new US speed. It is not the new Chinese speed.

Sources:

For longtime readers:

(2) Meta’s Zuck Veers Back to Open Weights

MP TAKE: Zuck is being pragmatic as ever, and notice the speed at which he is doing it: slowly for a change. Meta has gone open, then partly closed, and now kinda open again, one deliberate turn of the dial at a time. Each turn adapting to US policy winds and the competitive dynamics of the moment. That slow and steady steering IS the strategy.

Remember the arc. Until China’s DeepSeek showed up, early 2025, Meta was the global open source AI leader with Llama, throwing ‘sand in the gears’ of the frontier model companies, as I wrote back in 2024. Then Zuck veered closed, spending hundreds of billions on talent and data centers, and putting Alexandr Wang, acquired with his Scale AI in a $14.3 billion deal, in charge of Meta Superintelligence. The Muse generation went more closed than open.

Now the veer back, per the WSJ and Bloomberg: Muse Glimmer, an open weight 30 billion parameter model small enough to run on devices, think smart glasses and smartphones. And more notably, the flagship Muse Spark 1.2 is likely coming in an open weight version soon, returning Meta’s best to the Llama lineage.

Zuck wrapped it in a 6,500 word essay, ‘The Future is for Everyone’, his biggest AI statement in years. Writing essays is the thing among founder CEOs now, from Dario Amodei to Sam Altman, as I covered in ‘A Tale of Two AI Essays’, and lately Satya Nadella. Zuck’s argument: AI’s biggest risk is one entity with too much control, the closed labs’ discourse is ‘so filled with doom’, and you can learn from anything you can observe, a defense of distillation itself.

The chessboard note: this puts Meta squarely on the side of Jensen Huang’s 37-plus partner open secure AI coalition, which Meta joined, and the open-champion path we have argued Nvidia and Apple should lead. Watch for Elon to follow with Grok soon, and join the sand throwing, for the same reasons.

Sources:

For longtime readers:

(3) Microsoft’s Maia Makes the Nvidia-Alternative Chip Race

MP TAKE: Microsoft stuck to it. That is the whole story, and it is a big one. They used their market heft to will Maia into the cloud silicon race, and the standings now read: Amazon AWS at number one with Trainium, Microsoft Azure at number two finally filling in the middle with Maia, and Google Cloud at number three with its hugely successful TPUs. Slow and steady got Maia into the race.

The Information reports Microsoft’s homegrown AI chip effort is showing signs of life after a famously slow start, with the latest Maia versions finally good enough for business customers to kick around. Trainium took Amazon two to three years of investing to get interesting to businesses, including the frontier labs OpenAI and Anthropic. Google’s TPUs matured years earlier. I profiled the whole ‘Frenemies’ dynamic in Nvidia vs its Top Customers, and Microsoft’s silicon determination in its ‘Post-OpenAI’ path.

The stakes run through Taiwan. All three now contend for precious fab capacity at TSMC, which I think of as the ‘Fed of the AI world’, because they gate how many chips ultimately get made. Nvidia and Apple alone take over half of TSMC’s capacity, with Nvidia now TSMC’s biggest customer, a crown Apple wore a few years ago.

And before anyone gets excited about all this competition: Nvidia holds 80 to 90 percent global share of these chips, and that is not changing anytime soon. No less a personage than Elon Musk testified at his SpaceXAI earnings call last week that Nvidia’s Vera Rubin chips are the best, bar none. There is a real window for second sources in this supply-constrained environment, which runs through the end of the decade across GPUs, networking and memory. But the window for gaining ground narrows in three or four years. Everyone wants a second source. That is why the billions keep flowing.

Sources:

For longtime readers:

My overall take: all of these companies, software and hardware, are running the same slow and steady race, for different reasons.

Policy speed: cybersecurity pauses are becoming de facto US government policy, now increasingly bipartisan, from the White House framework to Sanders’ letter.

Strategy speed: the open versus closed dial, turned slowly and deliberately, is how Zuck adapts to Washington and throws sand at closed rivals in one motion.

Physics speed: chips take years and billions no matter who you are. Patience is not a virtue in silicon, it is the entry fee.

China is very much in the race with its own, different speed bumps on its own AI Autobahn: memory shortages, export controls, its own chip crunch. But the US speed limits are particular, and they are being debated in public, letter by letter, pause by pause. Remember how the fable actually works: slow and steady wins the race only when the hare naps. Nobody in Beijing is napping.

GADGET AI

OpenAI’s 2027 AI Device: A $300 Hockey Puck with Personality

MP TAKE: Look past the shape and you see the strategy. A displayless speaker with moving parts that give it ‘personality’ is a bet that the next AI device is a presence, not a screen. The puck is the wedge, not the destination. The real product is the ChatGPT subscription and the always-on relationship it rides on.

Bloomberg has the details: OpenAI’s first device, slated for 2027 and designed with Jony Ive, is a hockey-puck-sized speaker with personality, likely $300-plus, no display, with a battery so it moves from room to room. Essentially a super-sized ChatGPT optimized around a speaker, with subscriptions and profiles for the family. We covered the longer OpenAI hardware arc in the 2027 OS-driven AI smartphone ambition and the Jony Ive device challenges.

It enters a crowded, revamping field: Amazon’s Alexa Plus, Apple’s coming Siri AI and updated home speakers, Google’s Nest speakers with Gemini. But because it is OpenAI, and because it is Jony Ive, it gets the attention.

Sources:

For longtime readers:

Q1. What is most interesting to me about the OpenAI speaker?

The room-to-room mobility focus, and it is a curious one. Every other home speaker CAN be moved, but none is designed to be moved: you plug it into a corner and leave it. This one is built to travel the house with you, battery and all. Presumably OpenAI and Jony expect most people to buy just one and carry it around.

But here is the thing: the ultimate mobile, room-to-room AI device already exists, and it is in your pocket. My view: for the next two or three years, until something truly magical shows up from OpenAI or others, Apple runs away with the home and work device game. Siri AI, training on your personal data with trust, safety and security, will be one of the most personalized AIs out there, something no cloud-based service can match at that depth today, because they do not have your personal data to the extent Apple does.

Q2. What is the biggest headwind for OpenAI’s speaker?

Several, from distribution to making them at volume. But the biggest is ‘RAMageddon’, which even Apple faces: Apple just cited memory costs up about 40 percent on its upcoming devices, to be absorbed in margins or passed on in pricing. You already saw Apple TV, barely an AI device, go from $129 to $199.

So the $300 price being talked about likely lands closer to $500 to $600. This is the huge issue for local AI devices of all types, PCs, laptops, smartphones, and home speakers are no different. Very cool for the early adopters and gadget nerds who will crawl through glass to try OpenAI’s latest. Not a mainstream device yet. That is challenge number one.

WRAP

Today’s AI-RTZ #1174 is part 2 of the compute price series: The ‘Dwarkesh’ Compute Demand Bull Case, the strongest bull case fairly stated, and my ten brakes on it.

And a spoiler for tonight: AI-RTZ #1175 takes ‘RAMageddon’ to Washington, the memory lobbying wars, Apple testing Chinese CXMT chips, and the chessboard ahead of the September 24 Xi-Trump White House meeting. It foots directly with today’s show. Out in the wee hours.

AI Ramblings Daily on AI-RTZ is here to think through this AI Tech Wave, one day at a time.

Tomorrow, ARD 138 and AI-RTZ #1175.

Thanks for joining us, AI Curious Folk. Stay tuned.

— MP

Full Source Reading

For the broader context, see the canonical sources for today’s events.

Event 1. OpenAI’s Pauses

Event 2. Meta’s Open Weights Veer

Event 3. Microsoft’s Maia

Gadget AI. OpenAI’s 2027 Device

Clips from today

Clip 1. Meta’s Zuck Returns to Open Weights

Until China’s DeepSeek, Meta was the global open source AI leader with Llama, throwing sand in the gears of the frontier model companies. Then Zuck veered closed. Now he is veering back, with Muse Glimmer open weight and Muse Spark 1.2 likely to follow.

MP Take: Essays are the thing among founder CEOs, and Mark’s 6,500 word missive leans into Jensen Huang’s open secure AI coalition, Meta included. One deliberate turn of the dial at a time.

Clip 2. Apple’s Memory Costs Surge 40%

The biggest headwind for every AI gadget, including OpenAI’s coming device: ‘RAMageddon’. Apple just cited memory costs up about 40 percent on upcoming devices. Apple TV already went from $129 to $199.

MP Take: The $300 price talked about for OpenAI’s speaker likely lands closer to $500 to $600. Cool for early adopters who will crawl through glass for it. Not mainstream yet.

Clip 3. Microsoft’s Maia: New Player in AI Chip Race

The three cloud giants each have their own AI chips now: AWS Trainium, Azure Maia, Google Cloud TPUs, all contending for fab capacity at TSMC, the ‘Fed of the AI world’.

MP Take: Nvidia still holds 80 to 90 percent share, and Elon just called Vera Rubin the best chips bar none. The second-source window is real, and it narrows in three or four years.

Clip 4. AI Race: US & China’s Slow & Steady Approach

The US AI industry drove the Autobahn with no speed limits for three years. Now come the speed bumps, slow zones, and maybe speed limits, from cybersecurity pauses to bipartisan calls to slow down.

MP Take: All these companies are in the same slow and steady race now. China is in it too, with different speed bumps. The US ones are particular: cybersecurity and geopolitics.

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





Want the latest?

Sign up for Michael Parekh's Newsletter below:


Subscribe Here

More like this

$PTON Tumbles To New Low
$PTON Tumbles To New Low
Riley Rosebee Trends with No Friends
Good evening. Stocks stumbled. Every major index tumbled more than -1.0%. Every sector settled lower. Tech, healthcare and communications corrected more...
2 years ago
2 MIN READ
BROWSE ALL ARTICLES