levan retweeted
$MU $SNDK $SKHY $NBIS
Nomura’s memory team just put out a really interesting note on the AI scare, the numbers underneath are some of the strongest I’ve seen in this drawdown.
Nomura sees AI going from 10GW of data center power in 2025 to 40GW by 2028, taking it from 17% of all data center power to 35%.
The biggest slice of that 2028 number is heavy agents at 15GW, ahead of training at 10GW. In Nomura’s model, agents become the largest power user in AI within two years.
US big tech alone goes from 10.2GW of AI capacity to 48.5GW over the same stretch, Nomura’s memory team just put out a note on the AI scare, and once you get past the headline worry, the numbers underneath are some of the strongest I’ve seen in this drawdown.
Nomura sees AI going from 10GW of data center power in 2025 to 40GW by 2028, taking it from 17% of all data center power to 35%.
The biggest slice of that 2028 number is heavy agents at 15GW, ahead of training at 10GW. In Nomura’s model, agents become the largest power user in AI within two years.
US big tech alone goes from 10.2GW of AI capacity to 48.5GW over the same stretch, almost 5x.
The cash is piling up where the chips are made. Nine semiconductor companies (Samsung, Hynix, Micron, Sandisk, Kioxia, TSMC, Nvidia, AMD and Broadcom) went from $27B of net cash in 1Q24 to $282B in 2Q26, and Nomura projects ~$2T by end-2028.
The clouds buying all of it are printing too, with AWS at a 39% operating margin, Microsoft Intelligent Cloud at 41% and Google Cloud at 36%. Even Anthropic moved into adjusted operating profit in Q2, per Nomura.
Memory stays supply constrained through at least 2028. Nomura’s targets are 670,000 KRW for Samsung (+207%) and 4.7M KRW for SK Hynix (+268%).
Now the part the market sold on and I don't buy into:
OpenAI and Anthropic’s combined revenue grew 49% q/q in Q3, down from 74% in Q2. At September’s run rate, Q4 lands at $29.7B and 2026 comes in short of the ~$100B the market expected.
That “slowdown” scenario is still +291% y/y.
Part of the dip is mechanical.
OpenAI’s run rate got reframed because of how partner revenue is booked, Anthropic cut off distillation (5-7% of revenue), and summer slowed usage-based billing.
Meta moving engineers off Claude Code onto its own tools sounds bad until you remember the tokens still get burned, just on Meta’s compute. OpenAI and Anthropic are only ~5GW of the ~29GW of AI capacity Nomura sees by end-2026.
The 7 big operators, CoreWeave and Nebius included, go deeper into net debt through 2028. That debt is the chipmakers’ revenue.
CoreWeave, the name everyone is scared of, is 4.4% of capacity and 1.7% of 2026 data center investment. Nomura sees limited systemic risk and calls the CXMT memory panic excessive.
Every bearish line in this note ends up as someone’s GPU, HBM or optics order.
So invest in the entire buildout and win either way.
Bullish ⚡️
My unpopular opinion is that experience is now holding devs back.
We "know" what can't be ported, combined or rebuilt.
So we never try.
People who don't know the rules just try it.
And with Opus 5.5, it often works.
Try first. Ask why not later.
levan retweeted
traffic from ChatGPT is skyrocketing on my apps
and Ilias shares a gem here
this is such a good framework
side note: i looked into it and blog posts are 73% of what AI looks at when replying
blogging is faaarr from being dead
levan retweeted
This is probably the most important episode we recorded this year.
I'm not kidding, it's about everything:
How we work, why the product is the way it is, why we only have 21 people, why we still make a ton of money, and how we ship so fast.
21 people, a single meeting a week, no quarterly goals, no pull requests, no mandatory code review, no titles (except one dictator), a hundred releases a day, bugs fixed in minutes, famous for killing features.
@sqs and @thorstenball on how the Amp team works:
levan retweeted
21 people, a single meeting a week, no quarterly goals, no pull requests, no mandatory code review, no titles (except one dictator), a hundred releases a day, bugs fixed in minutes, famous for killing features.
@sqs and @thorstenball on how the Amp team works:
levan retweeted
Over 1 million lines of dead code removed. 🧹
Meet AutoClean, Uber's AI-powered system that's cleaned up 16,000+ stale feature flags using AI agents, deterministic code transformations, and validation.
levan retweeted
my unpopular opinion is that ChatGPT apps are the best distribution play rn
we spent last month shipping these for agency clients and...
+50-200% on income those apps were already doing
just you literally get an access to ~900M weekly users for free
+ you're not struggled with onboarding, since it's built on top of ChatGPT and for user who never used your product, much easier to get in via gpt interface
same product, same pricing, just new way of using your product
and the fear that OpenAI just clones you is backwards
cloning your niche is the ONE move that costs them more than it returns...
they'd rather buy you and ship it out of the box
here's exactly how to launch it:
1. pick a workflow the client already gets paid for
> not a new idea
> an existing paid workflow with proven demand
2. build it as an MCP server
> streamable HTTP on public HTTPS, usually at /mcp
> that server is the whole product surface
> UI is optional, tools come first
3. name and describe every tool like the model is reading it, because it is
> action-oriented name, a human title, a description of when to use it
> accurate annotations and a real input schema
> that metadata is what decides whether you get called at all
4. add the UI after the tools already work
> expose it as an MCP resource with a ui:// uri
> set the type to text/html;profile=mcp-app
> your built js gets inlined into that resource and runs in an iframe
5. connect it at chatgpt.com/plugins and break it yourself
> plus button, add custom MCP server
> paste the url with /mcp, create as a plugin
> run MCP Inspector first and catch the schema errors there
6. submit at platform.openai.com/plugins and overprepare
> a zip of the built plugin and publisher verification
> a privacy policy url, no privacy policy means no review
> 5 positive and 3 negative test cases
> release notes and a demo video for the reviewer
> a domain token hosted at /.well-known/openai-apps-challenge
7. keep billing on YOUR backend
> OAuth 2.1 into the account system you already have
> enforce authorization on every request, not through the model
> there is no revenue share
> the plugin is the channel and the MRR stays yours
that's absolutely free traffic and by some reason most of products skip it...
the guy below is a good case which shows real efficiency of this approach (he literally made around +$40k monthly income)
levan retweeted
I’m the CEO of Product Hunt and see 100s of launches a week. My current thoughts on new infra begging for startups:
1) Voice AI - infra for voice agents is getting really good w @pipecat_ai @Vapi_AI @DeepgramAI @ElevenLabs et al. and yet I rarely see good voice tech in the real world.
2) Decision Models - Jev / @typesafeai leading the charge. It reminds me of the mech-building in Zelda TOTK - an explosion of possibilities!
3) Personal Agents - they work now. And they need tools, apps, MCPs.
4) Extensions (OpenAI). We all live in agentic interfaces now. It’s like a new platform. Someone needs to build a Slack replacement that lives in Codex, for example.
levan retweeted
i'm begging you:
- find 20 startups that raised a seed round 3 to 9 months ago
(ignore any raise that happened <3 months ago)
- pick the 5 that sell to the same kind of customer
- find their founders on X
- read their last 100 posts
- pull out every question someone asked them about the product
(skip the ones the founder replied to)
1. why do all this?
[bookmark this right now, you’ll forget this by tomorrow]
because the replies under a founder's posts are people telling you, in their own words, what they want.
- people asking for alternatives
- anyone annoyed about the price
- "does it do X yet?" type questions
- people who signed up, got stuck, and gave up
put all of it in a sheet.
that sheet is now a list of possible customers, the problem each one has, and the exact phrasing they use for it.
next problem:
2. how do these people end up noticing you?
- take the 10 questions that come up most
- make one useful piece of content for each
- leave your product out of it
- answer the question better than the startup did
- tag the startup when you post it. if it helps their users, their team reposts it and you borrow their audience
- if an alternative does something better, tag them.
- if your answer needs 2 or 3 products chained together, write that down. that combo is your tool in step 4
>important<
someone asking whether an AI tool can do a thing?
- record a 30-second clip of you doing it.
when the complaint is pricing..
- line up 5 alternatives and show what each one costs once you add everything up.
if a workflow keeps tripping people up..
- build a free template that handles it.
do this for 30 days.
3. then check which posts sent people to your profile.
- ignore likes
- more bookmarks than likes means people plan to come back. repeat that topic
- replies that ask a follow-up question count more
- anyone who went and used one of your templates
- and note which problems keep bringing in people you haven't seen before
drop the topics nobody bit on and keep going on the ones that pulled people in.
4. now build a free tool for the problem that came up most.
one small tool is enough for now. the full SaaS can wait.
- put the link in your bio
- drop it under conversations where it fits
- send it directly to the people who asked for exactly this
- ask users what they still can't do with it
now you've got:
a) an audience that cares about one specific problem
b) a list of people who have already said they have it
c) a free product they have a reason to try
d) feedback from real users on what to build next
5. now add a paid feature.
dm your 20 heaviest users first with the price and two lines on what it does.
> 5 say yes, build it.
> fewer than 5, the feature is wrong, the price is fine.
Goal:
- $10/month
- 100 paying users
- $1k MRR
that's the target.
the order is the whole trick: customers first, then what they wanted, then distribution built around their problems, and something to sell only at the end.
If you read till here, reply “I’m a reader” helps me understand if I should keep my content this long.
a like, repost or a reply also encourages me to keep posting. ✨
levan retweeted
Meta Muse is now recommending our client's product as the best option on the market, then pushing people to sign up (see image below)
it's even sh*t talking their biggest competitor
this is the future of AEO
here's how to get there first (with 11 optimizations):
1. update your pages "accessibility tree"
this is literally how Muse reads your website.
do these four things:
1. add names to all icon buttons. For example, label your three line hamburger menu: aria-label="Open menu"
2. add a