Founder, CEO, Chief Analyst @MoorInsStrat. Co-founder of @TheSixFiveMedia and @Signal_65. Ex-AMD CVP & Fellow, Compaq and AltaVista dot-com bomb participant.
Austin, TX
Joined May 2008
- Tweets73.1K
- Following4.8K
- Followers69.5K
- Likes74K
Pinned Tweet
Meta's consumer AI comeback looks real. The bill already is: last reported quarter, capex ate nearly every dollar its business generated. It's The Week Ahead In Technology for October 19, 2026.
Tuesday I'm in Washington at the @Google Public Sector Summit.
$META guided $130 to $145 billion of capex this year. Zuckerberg says it could sell compute at a big markup, but AI services earn higher margins. I want next year's number.
@NVIDIAGTC Berlin: Jensen Huang's keynote. Boardrooms want sovereign, secure data.
$SPCX, then $TSLA: I'm listening for chips and compute.
$IBM and $SAP: show me AI agents in booked revenue, not pilots.
$NOK: is the @nvidia deal paying off?
Carry-over: chip stocks fell eight days straight. I've said it's just another scare.
Can Meta's comeback pay the bill?
This week, chip sales cross a trillion dollars a year for the first time, and the transistor isn't what limits AI anymore. Power, memory and packaging are. It's The Week Ahead In Technology for October 12, 2026.
@OpenComputePrj Global Summit: @AMD's @LisaSu and NVIDIA keynote. The test is the whole system: 800-volt power, liquid cooling, networking and racks.
SEMICON West: @Applied4Tech hosts investors. Its plan to double capacity by 2028 is the loudest demand signal a toolmaker can send.
$ASML and $TSM report. On TSMC, I'm listening for allocation and packaging capacity.
Apple's "Welcome home" is widely reported to be smart home hardware. Apple already put a meter on Siri. I want to know if the meter comes home too.
So which limit breaks first: power, memory or packaging?
Duet, @bot, Dot, MuseBot here we go folks.
Claude Duet is an early research preview, available through a waitlist on Pro and Max plans in the US. It can still misjudge what’s urgent, and we’ll keep tuning it based on what early users tell us.
Join the waitlist: claude.com/product/claude-du…
Read more: claude.com/resources/article…
User error!
"Computer" has always been cloud-based even though it had the Computer name and Computer is an extra charge and I still use it for 4-5 company workflows.
To use Local, you choose "Computer"!
Thanks @themtsyh!
And @MicrosoftEdge also rewards my loyalty by asking me about Rewards even though I have clicked "Don't show this recommendation again" at least 100 times.
Around once a month, @MicrosoftEdge honors my loyalty by trying to switch me off of Google as the default search engine. Always makes me question how much Bing traffic is by accident or by choice.
Around once a month, @MicrosoftEdge honors my loyalty by trying to switch me off of Google as the default search engine. Always makes me question how much Bing traffic is by accident or by choice.
Day 3 @perplexity_ai Local inference still busted on my 128GB RTX Spark Laptop. How on earth do you get help from the company? Help @AravSrinivas I have spent tens of thousands of dollars with your company.
Brain-dead simple but can’t get it to work. No local model in the model pull down.
Yes, I’ve closed and restarted the app.
Yes, the model says it’s running in settings.
Yes, the app is the latest.
Help @perplexitydevs.
Daniel and I debate the value of today's benchmarks in the Flip. Serves vendors or serves buyers?
Are AI leaderboards helping enterprises choose the right models, or mainly serving vendors?
On The Flip, Ep. 324 of The Six Five Pod, @PatrickMoorhead argues for real-world benchmarks and cost per successful outcome. @danielnewmanUV defends independent rankings as a way to narrow the field before hands-on testing.
The takeaway: Cost per token isn't enough. Cost per known good output matters more.
The Flip assigns opposing positions that don't necessarily reflect the hosts' views.
Looking forward to chatting with @totobossf1 and @AntonioNeri_HPE for the F1 COTA! I probably shouldn't mention I'm an RB fan.
F1 meets enterprise tech in Austin!
Oct. 22, 2026, @PatrickMoorhead and @danielnewmanUV will be on their home turf at the @F1 US Grand Prix, speaking with @TotoWolff_ and @HPE CEO @AntonioNeri_HPE about their Mercedes-AMG PETRONAS F1 partnership and how AI, data, and advanced computing drive performance.
Stay tuned!
The new unit of SI/AI is now a pre-packaged, pre-validated pod. OEMs have been shipping pods for decades, and this is a reference design. Think of improved TTM and simplicity. The easy button. Partners include Foxconn, Gigabyte, Lenovo, HPE, Dell, Quanta and Pegatron.
AI factory deployment at a prepared site: from six months down to 14 days.
NVIDIA DSX POD brings Vera Rubin AI factories in containers, integrated and validated before delivery.
Announced at GTC Berlin: nvda.ws/4slQvNp
Patrick Moorhead retweeted
So last week, I published my review of the @NVIDIARTXSpark platform, today I am sharing my review of the two systems I tested with that platform, including the @Microsoft @Surface Laptop Ultra and the @HP OmniBook Ultra 16. Both laptops bring something different to the table for the Spark platform and try to accomplish different things with that much memory (128GB) and performance. That said, both still have room for improvement in some areas, but are both a great first try on a new platform.
moorinsightsstrategy.com/fie…
I walked right past Nick Shirley on the street today in DC. I didn’t recognize him as he was in a suit, not a hoodie. From far away, it looked like three DC interns until he got closer. I told him “keep up the good work”.
.@AFergusonFTC: "A lot of journalists in the room today — everyone knew about the Quality Learing Center... No one reported it... It took one brave kid and a camera to make everyone aware of this, and it launched a revolution."
Gemini 4 Argon…
Google's AI was written off. Now Gemini 4 Argon tops @Signal_65's Pinnacle composite, ahead of several GPT and Claude models. @danielnewmanUV and @PatrickMoorhead break down the performance of Meta's Muse Wonder, NVIDIA's fully open Nemotron, and why AI leadership keeps rotating. $GOOGL $META $NVDA
Karen Dahut, CEO of Google Public Sector, kicks off its Public Sector here in Washington, DC. The theme is “where mission meets moonshot”.
As you would expect, Google is leaning into the full stack that has been very successful in the past few years in the federal space.
Customers cited so far include NASA, NOAA, the Department of War, the Department of Energy, the U.S. Army, and America.gov.
Patrick Moorhead retweeted
Meta's consumer AI comeback looks real. The bill already is: last reported quarter, capex ate nearly every dollar its business generated. It's The Week Ahead In Technology for October 19, 2026.
Tuesday I'm in Washington at the @Google Public Sector Summit.
$META guided $130 to $145 billion of capex this year. Zuckerberg says it could sell compute at a big markup, but AI services earn higher margins. I want next year's number.
@NVIDIAGTC Berlin: Jensen Huang's keynote. Boardrooms want sovereign, secure data.
$SPCX, then $TSLA: I'm listening for chips and compute.
$IBM and $SAP: show me AI agents in booked revenue, not pilots.
$NOK: is the @nvidia deal paying off?
Carry-over: chip stocks fell eight days straight. I've said it's just another scare.
Can Meta's comeback pay the bill?
This week, chip sales cross a trillion dollars a year for the first time, and the transistor isn't what limits AI anymore. Power, memory and packaging are. It's The Week Ahead In Technology for October 12, 2026.
@OpenComputePrj Global Summit: @AMD's @LisaSu and NVIDIA keynote. The test is the whole system: 800-volt power, liquid cooling, networking and racks.
SEMICON West: @Applied4Tech hosts investors. Its plan to double capacity by 2028 is the loudest demand signal a toolmaker can send.
$ASML and $TSM report. On TSMC, I'm listening for allocation and packaging capacity.
Apple's "Welcome home" is widely reported to be smart home hardware. Apple already put a meter on Siri. I want to know if the meter comes home too.
So which limit breaks first: power, memory or packaging?
I’m speechless at how wrong this is. And when do we listen to Gartner on what’s happening with hyperscalers which are driving most of the demand.
“.. At the moment, we have more compute capacity than we have demand, and we have a lot of the tokens and processing power being given away in freemium solutions,” he told @CNBC. “Everyone in the AI race is running for a last-man standing position.”
cnbc.com/2026/10/19/gartner-…
Meta put out a deep dive today on “Prometheus”, its gigawatt-scale training cluster in New Albany, Ohio, and it's a great example of co-engineering to the max with @nvidia. Meta says it goes beyond a normal vendor relationship with teams embedded in the data centers, sharing real-time telemetry and running joint triage on every halt.
Hundreds of thousands of NVIDIA GPUs across multiple buildings today, roughly 1M at full scale, and phase one trained the @Muse models. I got an early read over the weekend. What I liked most is Meta put real failure data out there. You don't see that very often.
Here are the deets:
•in mid-2025, Meta was seeing nearly 1,000 interruptions per 8K GPUs per DAY. Down 99% since.
•the goal is 2 per 8K GPUs a day, so by my math they're still about 5x off.
•each 72-GPU rack runs a 64-GPU job slice because training likes powers of two. The other 8 are hot spares, plus ~2% extra racks. Roughly 1 in 8 GPUs in a big run sits in reserve.
•GB300 was 5-10% faster than GB200 for Meta. Looks like Meta pulled 800G per GPU forward on its GB200s, so it didn't get that bump. Meta expects Vera Rubin to add 10-20% perf/watt.
•Arista chassis on Broadcom Jericho3 connect buildings anywhere from 2 km to 100+ km apart.
•starting or pausing one big job can swing tens of megawatts in seconds.
At this scale you're fighting reliability and power every day, and that co-engineering matters as much as the chips.
Good fodder ahead of Meta earnings Wednesday.
$META $NVDA $ANET $AVGO
engineering.fb.com/2026/10/1…
