Takeaways: Server CPU Outlook
> Rumors of Meta and Microsoft cutting server CPU orders are overblown. Microsoft’s total order volume is unchanged—only manufacturing allocation across ODMs was reshuffled.
> Meta is trimming legacy x86 orders (primarily Intel Emerald Rapids) to aggressively scale Nvidia’s Arm-based Grace and Vera CPUs (>500k chips).
> Overall x86 server CPU shipments reached 8.13M units in 3Q26 (>20% YoY growth), marking the fastest expansion in roughly five years.
> AMD's upcoming EPYC Venice line is projected to ship 400k units in 2026 before surging to 5.4M–6M units in 2027.
> AMD hit a record 35.5% x86 server share (+1 percentage point QoQ, and up 7.7 percentage points from 27.8% in 3Q25), largely at Intel’s expense.
> The global server CPU TAM is still projected to hit $211B by 2030, representing a 47% CAGR.
> Total server CPU unit shipments are forecast to grow 29% in 2027, propelled by a projected 79% jump in AI inference-related CPU demand.
$NVDA $META $MSFT $AMD $INTC $ARM
Show more
Amazon $AMZN CEO on AI Compute:
"Big chunks of capacity are largely reserved already in 27, much of it is reserved already in 28, and a good chunk in 29" 😱🔥
"The economics of the business are such that [once monetized] you really like the FCF and ROI capital"
$NBIS
Show more
#
Nvidia#'s interim CEO says no unlimited spending on startups. Its biggest bets include OpenAI at $30B, Groq at $20B (license and acqui-hire, media-reported) and Anthropic at up to $10B. 💡More: 🔗 $NVDA
Show more
Banger paper from Google on self-improving agent harnesses.
(bookmark it)
It addresses common issues in self-improving agent harnesses like local overfitting, catastrophic forgetting, and memory pollution.
They propose GUARD, which keeps the base model frozen and lets the agent edit its own prompts, skills, tool rules, and memory. An edit is accepted only when its gains outweigh the behaviors it breaks.
- On 43 hosted SDK tasks, GUARD reached 80.80% accuracy, 8.89 points over the starting harness (p = 0.008).
- It beat Meta-Harness by 6.28 points and used 54.5% fewer tokens per task.
- On procedural memory tasks, it led update correctness at 90.1% with zero memory pollution. Aggressive rewriters scored similarly and tripled pollution.
How does it know what broke? Each task is split into self-checks before execution. Every round grades trajectories against them, pins rules the agent already handles, and ranks newly broken behaviors first. At test time, one frozen harness runs in a single pass.
Chat with Paper:
Show more
Going back to $AAOI Q3 2026 earnings. Why did the company disappoint with its Q4 guidance?
$AAOI guided for $385–420M in Q4 revenue, below market expectations.
The reason is quite interesting.
One of AAOI's major hyperscaler customers is transitioning from 800G to 1.6T transceivers faster than management anticipated.
As a result, AAOI expects to sell fewer 800G transceivers than originally planned, but it can't ramp up 1.6T production quickly enough to offset the difference.
Why? They didn't secure enough 1.6T DSPs in advance.
AAOI currently has two qualified DSP suppliers (I assume Marvell and Broadcom) and is working on qualifying a third (my guess is Credo).
These are my assumptions, as management hasn't disclosed the suppliers' names.
According to management, DSP availability should improve significantly in Q1 2027.
If the third supplier turns out to be $CRDO, the timing would be interesting, considering Credo is also expecting a significant ramp in its optical business during FY2027.
However, this is just a hypothesis, not a confirmed relationship between the two companies.
I also think the recently announced ATM could partly be related to securing the supply chain.
AAOI needs significant capital to support its manufacturing expansion. It's not just about building factories and purchasing equipment. They also need to secure DSPs, wafers, substrates and other components months in advance.
And this isn't just an AAOI problem. Other optical transceiver manufacturers have also mentioned supply constraints and indicated they could produce more if certain components were more readily available.
Another thing I'll be watching closely is inventory.
AAOI's inventory increased significantly in Q3, reaching approximately $378M.
I expect another substantial increase in Q4 as the company prepares for higher production volumes.
For me, an increase in raw materials and components would be a positive signal that AAOI is securing its supply chain ahead of the planned ramp.
Of course, inventory growth alone isn't enough. Ultimately, I want to see those components turn into finished products and revenue.
But what does the faster transition from 800G to 1.6T actually mean for AAOI?
This is probably the most important part.
It means the company's product mix is shifting toward higher-priced transceivers faster than originally expected.
And that creates another opportunity for revenue growth.
According to industry pricing estimates, a 1.6T transceiver can cost roughly twice as much as a comparable 800G transceiver, depending on the configuration.
So AAOI has two potential drivers of revenue growth:
Increasing the number of transceivers it can manufacture.
Increasing the average selling price as 1.6T becomes a larger share of total shipments.
In other words, AAOI doesn't necessarily need to double its production volume to double its transceiver revenue, assuming the product mix shifts toward higher-priced products.
And I expect this mix to continue shifting toward 1.6T as hyperscalers upgrade their networks.
What's interesting is that the same transition hurting AAOI's Q4 revenue could become one of its biggest growth drivers in 2027.
Of course, management still needs to secure the components, complete supplier qualifications and deliver the production ramp.
But I see the faster transition to 1.6T as a positive signal for long-term demand, even if AAOI wasn't fully prepared for it in Q4.
Show more
That’s the signal of confidence I like to see: $COHR issues a new $2B share buyback, through open market purchases.
Optical industry looks very undervalued right now, especially as it reaches the inflection of the next supercycle.
We’re already seeing photonics-SOI substrate players go from $25M revenue -> $1.2B LTAs signed in a quarter.
Or $100M backlog -> $500M in a quarter with InP substrates.
And we’ll see acceleration for lasers, modules, and the optical supply chain next.
I’m sure Coherent finds its own valuation very low now too, especially after signing $20B LTAs recently.
Show more
Only 0.1% left for 4k followers 👀✨️