Arm Holdings Just Changed Its 35-Year Business Model — From Licensing to Direct Sales
Arm Holdings Just Changed Its 35-Year Business Model — From Licensing to Direct Sales
Arm Just Changed Its 35-Year Business Model
On the surface, Arm Holdings' March announcement looked like a product lineup expansion. Look closer and it's a 35-year-old company quietly changing what kind of business it is.
Arm doesn't make chips. That was the identity. They license chip designs to Apple, Qualcomm, Samsung, Nvidia — every one of those companies pays Arm a royalty on every chip they ship. That's why their gross margins look like a software company, not a semiconductor company.
In March, that model changed. Arm is selling chips directly to customers for the first time ever.
The Core Announcement — Arm AGI CPU
The product is called the Arm AGI CPU. The first committed customers are Meta and OpenAI. Production is expected to start by the end of 2026.
This isn't just 'licensing plus direct sales' as a dual track. Arm reorganized into three new business units:
- Edge — devices (smartphones, embedded)
- Physical AI — robotics, autonomous systems
- Cloud AI — data centers
All three units are aligned with different layers of the AI cycle. And the Cloud AI side was already exploding quietly before this announcement.
Already the Hyperscaler Standard — Data Center Royalties Doubled
Here's the part people miss. Before the announcement, Arm was already the de facto standard inside data centers.
- AWS Graviton — Arm Neoverse
- Google Axion — Arm Neoverse
- Microsoft Cobalt — Arm Neoverse
- Nvidia Grace — Arm Neoverse
All four major hyperscalers run their proprietary data center chips on top of Arm's architecture. Put differently, Arm is already the standard ISA for cloud compute.
The numbers make it concrete:
- Data center royalties: doubled year-over-year
- Gross margin: 97%
You don't see 97% margins anywhere else in semiconductors. That number is possible because Arm doesn't manufacture — they license designs. Almost every dollar of revenue drops directly to operating income.
Why Pivot to Direct Sales Now
This is where it gets actually interesting. The royalty model is working. Why take the risk of going direct?
Two reasons.
First, the dollar value of cloud AI compute is rising fast. A per-chip royalty is on the order of cents to a few dollars. A custom data center chip sells for thousands. For the same underlying design, the difference between licensing it and selling it directly is orders of magnitude.
Second, customers like Meta and OpenAI want to buy direct. They don't have time to spin up their own silicon teams, and they don't love being entirely dependent on Nvidia for both supply and pricing. If Arm shows up with a 'ready-made AI-optimized CPU you can just buy,' there's a market.
Power Efficiency — Why Arm Actually Wins in the Data Center
One technical point. Arm-designed chips run cooler and use less electricity than competing architectures. The same property that matters in your phone is decisive in a data center.
We're at the point where hundreds of thousands of AI accelerators sit in a single building. Power and cooling now account for more than half of data center operating cost. Hitting the same performance at 30% less power is the margin.
That power advantage is the reason AWS, Google, Microsoft, and Nvidia all built their custom chips on top of Arm. Now Arm is taking that advantage and selling it directly inside their own silicon.
What to Watch
Three signals to track on this transition:
- First deployment outcomes from Meta and OpenAI. When production starts at end of 2026, the inflection point is whether they come back with follow-on orders. Fast re-orders mean Arm is now a direct competitor to Nvidia inside data centers, not just an underlying architecture.
- How existing licensees respond. Apple, Qualcomm, Samsung — their supplier just became a competitor in adjacent markets. If license renewal negotiations start dragging out, that's a signal.
- Whether data center royalty growth is sustained. Doubling YoY can be a quarterly event or a multi-quarter trend. The next two prints answer that.
The biggest bet I see here is simple. Arm already dominated mobile. They're now positioning to dominate data centers. And they're doing it while keeping the royalty business running — and entering the highest-margin new market directly. Transitions this clean don't happen often.
More in this Category
Coherent Wins — A Six-Round Scorecard for Five AI Infrastructure Stocks
Coherent Wins — A Six-Round Scorecard for Five AI Infrastructure Stocks
I scored Coherent (COHR), CoreWeave (CRWV), Nebius (NBIS), Iren (IREN), and Applied Digital (APLD) across six rounds. Coherent took it with 10 points, driven by the only debt-to-equity ratio under 32%.
Five Rules for Treating AI Infrastructure Stocks as Tactical, Not Core
Five Rules for Treating AI Infrastructure Stocks as Tactical, Not Core
Debt-to-equity across the five AI infrastructure plays spans 31% (Coherent) to 387% (CoreWeave). Here are five rules I use to treat them as tactical trades, not core holds.
Memory Sold Out Through 2027: Why Micron Now Prices Like a Utility
Memory Sold Out Through 2027: Why Micron Now Prices Like a Utility
Micron's HBM lines are effectively sold out through 2027, and that supply-demand gap is flowing directly into quarterly margins. Why waiting for a $480 pullback beats chasing the vertical line, and the three scenarios that would actually break the thesis.
Next Posts
When Fear Peaks, Fortunes Are Made: What 1929 Through 2020 Actually Taught Us
When Fear Peaks, Fortunes Are Made: What 1929 Through 2020 Actually Taught Us
Five historic crashes from the Great Depression to the 2020 pandemic show the same pattern — investors who bought at peak fear compounded at 11.5% to 20% annually for decades.
The Four Rules That Separate Real Crash Buyers From Pretenders
The Four Rules That Separate Real Crash Buyers From Pretenders
Buffett and Munger's framework for actually buying when markets collapse — a prepared watchlist, dry powder, emotional discipline, and accepting you won't catch the exact bottom.
Buffett Indicator at 137% — Today's Market Is More Expensive Than The Dot-Com Peak
Buffett Indicator at 137% — Today's Market Is More Expensive Than The Dot-Com Peak
Market cap to GDP sits at 137% overvalued versus 47% at the 2000 top, with CAPE at 42.05. Don't stop dollar cost averaging, but prepare for 8 to 15 years of choppy returns.
Previous Posts
Buy-and-Hold Is Dead: What -85% to -99% Crashes Tell Us About 2026
Buy-and-Hold Is Dead: What -85% to -99% Crashes Tell Us About 2026
Over the past 12 months PayPal fell 85%, Rivian 92%, Beyond Meat 99.7%. Here's why "buy a famous company and bury it" no longer works in 2026 — and what changed.
Wall Street Doesn't Hold Anymore: How to Follow the Money's Footprints
Wall Street Doesn't Hold Anymore: How to Follow the Money's Footprints
Institutions moving hundreds of billions of dollars can't hide. Here's how to read the footprints they leave in price action, volume, and sector ETF relative strength.
Single Stock vs Basket Investing: Which Survives 2026?
Single Stock vs Basket Investing: Which Survives 2026?
In the AI era, guessing which company wins is far harder than guessing which industry rises. Here's a side-by-side of basket investing vs single-name bets — and when each makes sense.