# data center revenue — X 热门讨论 (2026-10-02 12:36 UTC)
## @Funmentalist (The Fundamentalist) · 10-02 07:13 · ♥44 ↻1 💬1 $NBIS Inferize Acquisition Impact on Margins Analysis
This got way more complicated than I thought it would be, so let me share how I approached this analysis.
First, we have to start with the current margin and move from there. Nebius AI's business generated a 50% adjusted EBITDA margin, with management explicitly saying that revenue from Token Factory is "high-margin".
With today's announcement of Inferize, both capacity utilization and token economics should improve, so let’s do a little stress-test to understand what could theoretically happen to Token Factory’s margins under certain assumptions:
1) The model architecture
The economic impact Inferize should provide goes two ways:
Inferize benefit = cold-start recovery + reduction in warm spare GPUs
The cold start is obvious: when you need to launch another model replica and load more GPUs, you are losing time on that cold start. But there is one more thing that might be even more important and that is the reduction in warm spare GPUs.
Simple example:
Imagine Token Factory normally needs 800 GPUs, but traffic can suddenly spike and require 1,000 GPUs. This is, however, a problem because if launching another model replica takes 23 minutes, Nebius can’t wait until demand appears and then start those 200 GPUs. It would cause issues such as slowdowns or even failed requests for the customer.
So Nebius might keep extra model replicas already running, resulting in:
800 GPUs serving normal demand + 200 GPUs already loaded and waiting
Nebius even says that traditional cold starts force platforms to hold spare capacity to meet service-level targets. Inferize should allow Nebius and their capacity to scale much more closely with the actual usage. How?
If now launching capacity takes 20 seconds and not the current 23 minutes, Nebius may need only 50 warm GPUs waiting because they know that they can launch these additional 150 GPUs in 20 seconds, very important detail that people might miss.
This implies the following effective uplift:
Token Factory effective capacity uplift = cold-start recovery + warm spare recovery
Then:
Revenue uplift = capacity uplift x monetization rate
Which, in the end, causes:
EBITDA margin = (M_0 + u x C) / (1 + u)
I know the formula looks complicated, but realistically it is just:
- M_0 = existing Token Factory margin
- u = additional revenue from Inferize
- C = incremental contribution margin on that revenue
2) Cold-start math
I will keep on using the @EndicottInvests example that I commented a bit in different post, so based on that benchmark, we should expect this shrinkage of duration:
23 minutes -> 20 seconds (est. saved time: 22.67 minutes)
For every GPU repeatedly exposed to this:
capacity recovered = (22.67 x cold starts/day) / 1440
As I discussed in my previous post, we have to stress-test and try different inputs into our model because we don’t have a single number that would tell us average number of cold starts per day, so we get this:
Cold starts per affected GPU/day -> capacity recovered
0.25 -> 0.39% 0.50 -> 0.79% 1.0 -> 1.57% 2.0 -> 3.15% 4.0 -> 6.30%
3) Realization of theoretical savings from cold-starts
We shouldn’t expect that every theoretical saved minute will become usable commercial capacity as there can be still factors such as scheduling, networking delays, orchestration overhead, and other software bottlenecks (way above my paygrade):
So let’s assume 3 realization percentages:
60% 80% 90%
4) Warm spare capacity freed
As explained above, without long cold-starts, Nebius might not need that many extra GPUs ready for handling unexpected demand spikes.
Probably the biggest variable, so the stress-test requires a bit wider range of possibilities and we will use:
2% 5% 8% 11% 15% 20%
Just for imagination, 8% means that this percentage of Token Factory GPU capacity can become productive instead of sitting warm/not utilized.
5) Monetization of recovered capacity
I wouldn’t be worried about Nebius not being able to sell additional available capacity because the current demand is absolutely mindblowing. However, we shouldn’t assume that every GPU-hour recovered immediately turns into revenue.
Some capacity might remain temporarily unused, some might be allocated for redundancy or whatever.
So we will stress-test with this:
70% 85% 100%
6) Incremental contribution margin
Now we are getting into the interesting part and that is margin, because obviously the new incremental revenue should have higher margin than Token Factory’s existing average margin as GPUs are already purchassed, data center capacity exists, networking infrastructure is installed and engineering team is already employed.
But we still have some incremental costs such as electricity, networking, usage, etc.
Let’s assume the following margins:
70% 80% 90%
For clarity, this is not Token Factory margin, but margin from additional revenue created by better utilization of the already-installed infrastructure!!!
7) Token Factory margin
Nebius does not disclose Token Factory EBITDA margins, but if we know that Nebius AI generated ~50% EBITDA margin and Token Factory revenue is high-margin, let’s use these stress-test margins:
55% 65% 75%
8) Total capacity uplift
Now we can combine the first two operating effects:
Total capacity uplift = realized cold-start recovery + warm spare capacity freed
Example:
1 cold start/day = 1.57% theoretical recovery
with 80% realization: 1.57% x 80% = 1.26%
add 8% of warm spare capacity freed thanks to Inferize:
1.26% + 8% = 9.26% of effective capacity uplift
apply the monetization rate of 85%:
9.26% x 85% = 7.87% revenue uplift
So Token Factory could theoretically (under assumptions in this example) generate roughly 8% more revenue from the same installed GPU base as before.
9) Revenue uplift -> margin expansion
Back to the original formula for margin:
EBITDA margin = (M_0 + u x C) / (1 + u)
Example:
Token Factory now:
Revenue - $1B $650M of EBITDA EBITDA margin - 65%
If we use our revenue uplift of 8% from the same installed infrastructure and we assume that this additional revenue has 80% margin:
$80M of additional revenue $64M of incremental EBITDA
Token Factory with Inferize:
Revenue $1.08B $714M of EBITDA EBITDA margin ~ 66.11%
so the improvement of EBITDA margin is 1.11% percentage points.
But we use a lot of assumptions and everyone might have a different view on them and also the Token Factory environment might change quickly in next months, so this is why we stress-test.
Overview of stress-test
The full model assumes six main variables with different inputs:
1) Cold starts per day
0.25 0.5 1.0 2.0 4.0
2) Cold-start realization
60% 80% 90%
3) Warm spare capacity freed
2% 5% 8% 11% 15% 20%
4) Monetization of recovered capacity
70% 85% 100%
5) Incremental contribution margin
70% 80% 90%
6) Starting Token Factory margin
55% 65% 75%
This creates a very wide range of potential options and combinations, so I decided to build a complete solution with all the matrixes you need to look at all possible combinations. You can access the file on my Substack here: https://t.co/UnWVkSgq2d
For people interested in my opinion, in the picture you see my current base case, which I think is pretty significant:
1) 6.66% revenue uplift 2) 1.56% EBITDA margin uplift https://x.com/Funmentalist/status/2105918979042611597
## @PhotonCap (Photon Capital) · 10-01 22:45 · ♥30 ↻6 💬1 The Silicon Photonics Foundry Layer: Who Can Actually Make a PIC https://x.com/PhotonCap/status/2105791194173526473
## @HunterAllen4 (THE GAP FATHER) · 10-01 14:22 · ♥30 ↻0 💬12 $AXTI
is sitting in one of the cleaner setups on my screen right now.
The stock has repaired hard off the September lows and is compressing directly underneath the $79.80–$80 breakout zone.
The setup is simple: I want to see a clean break, acceptance above $80, and ideally a controlled retest that gives defined risk. I’m not interested in chasing an extended candle.
But the chart is only half the story. AXT is essentially the export valve for a critical piece of the optical AI supply chain. AXT is headquartered in the U.S., but its InP crystals are grown through Beijing Tongmei in China. Since February 2025, Chinese exports of InP, GaAs and germanium substrates require Ministry of Commerce permits.
That creates a situation where the constraint isn’t necessarily finding customers it is getting qualified wafers out of China and into the hands of the customers who need them.
Demand is already showing up. AXT generated $47.6M of Q2 revenue, up 165% YoY, with InP contributing roughly $30.7M, up 284%. Data-center InP hit a record quarter.
Management has been guiding manufacturing capacity toward roughly $60M per quarter by year-end 2026 and approximately $130M per quarter by year-end 2027, with 6-inch InP becoming increasingly important. Backlog is already above $100M and extends into 2027. The customers are knocking; the question is how quickly AXT can turn crystal capacity and permits into shipments.
Then you have the optical demand downstream. Lumentum has said InP laser demand is running more than 30% ahead of what it can currently ship, while $COHR Coherent has also described InP as a major industry constraint. That’s where the Oracle/Lumentum optical-circuit-switch story becomes interesting.
The OCS itself isn’t an InP product it’s a MEMS switching fabric so Oracle discussing production-scale OCS deployment wouldn’t suddenly create an AXT order.
But the optical links surrounding that switching infrastructure still need InP-based EMLs, continuous-wave lasers and increasingly high-power external lasers for CPO. More AI scale-out means more optical infrastructure, and that means more substrate underneath the lasers.
AXT also has meaningful demand visibility already locked in. Lumentum has a capacity reservation through December 31, 2031, including a $43.5M deposit now and another $43.5M scheduled for 2028.
Coherent has a separate multi-year 6-inch InP development and supply agreement backed by a $22.3M prepayment covering 2026–2028. Casela also has committed volume extending into 2027.
The China competition is real, but it doesn’t instantly solve the problem. Yunnan Germanium, Zhuhai Dingtai Xinyuan, Vital Materials and others are expanding InP capacity, while Sumitomo and JX are adding Japanese capacity. The issue is qualification, exportability and timing.
New crystal capacity can take 12–24 months to qualify, and Chinese material going overseas still faces the same licensing framework.
The biggest risk is therefore structural: AXT’s China manufacturing footprint makes permits a major variable.
Management has said permits remain its most significant challenge, and U.S.-bound InP shipments had still not been fully cleared as of the August disclosure. Europe and Japan have seen permits reopen, while China domestic demand can be served without the export bottleneck.
If U.S. permits remain stuck, Lumentum and Coherent commitments could take longer to translate into recognized revenue. If permits continue opening, AXT’s capacity expansion gives it a much larger runway into 2027.
The October 15 Oracle/Lumentum OCS discussion is useful confirmation for the broader scale-out story, but the cleaner AXT signal is the external-laser/CPO discussion and then Lumentum’s November earnings.
If $AXTI takes $80 and holds it, this one could wake up very quickly.
100$ will come quick. Don’t sleep 😴 https://x.com/HunterAllen4/status/2105664640306774142