Arm stops licensing, starts shipping: its own 136-core cpu is meta’s new ai engine
Arm just tore up its 35-year-old playbook. Instead of collecting royalties on other people’s silicon, the British company will sell a 136-core, 300-watt CPU it designed, branded and priced itself. The stock popped 10% before the bell in New York, adding $12 billion in market cap before traders had finished their coffee.
Meta signs first, queues form
Meta is the anchor customer. The social-network-turned-ai-farm will drop the unnamed chip—internally tagged CPU AGI—into its next-generation training clusters, replacing chunks of the Intel Xeon and AMD Epyc racks it currently leans on. Santosh Janardhan, Meta’s VP of infrastructure, calls the move a “step-function jump in performance per rack,” a phrase that usually precedes very large purchase orders.
Arm CEO René Haas doesn’t bother with modesty: “We have customers queuing to buy it.” Translation: hyperscalers have already forecast demand in the millions of units, enough to justify a $15 billion annual revenue line by 2029. That is more than triple Arm’s entire top line today.

The margin math flips overnight
Here is why Wall Street is rerating the stock. When Arm licenses its instruction set to, say, Qualcomm, it keeps roughly 5% of the chip’s selling price. When it licenses a full core design, the take rises to about 10%. When Arm becomes the merchant, the gross margin jumps to 50% on a $1,000 data-center CPU. Sell a million of those and you just minted half a billion in gross profit the old model never touched.
CFO Jason Child spelled it out for analysts: the licensing business will still grow—he pencils in $10 billion by 2029—but the new chip franchise will be bigger, riskier and vastly more lucrative. Total company revenue: $25 billion, five times today’s level. EPS: $9 against the $1.75 consensus for this fiscal year. The model isn’t changing; it’s being duplicated at a higher octave.

Tsmc fabs, nvidia competes, china beckons
Taiwan Semiconductor will build the processor on its 3-nm node, the same line that cranks out Apple’s M-series and Nvidia’s Blackwell GPUs. Arm will not own fabs, will not inventory wafers, will not touch capital-intensive depreciation. It becomes, overnight, a fabless competitor to AMD and Intel while still collecting checks from those very rivals for architectural licenses.
Haas swears his 136-core part is “complementary, not competitive” to Nvidia’s Grace CPU, but the spec sheet tells another story: identical TDP, comparable memory bandwidth, and a marketing deck that benchmarks Grace on power efficiency—then shows Arm ahead by 30%. Nvidia’s response is already in the wild: Grace Hopper Superchips shipping this quarter. The fight for the control plane of ai clusters just turned into a bar brawl.
China remains an open question. Haas says the core IP is British-born and therefore “unlikely to trip U.S. export controls,” a nuance that matters if Washington tightens the screw on ai hardware. One Shenzhen cloud provider has already asked for 50,000 units, according to a distributor note seen by Coastal Code. The deal isn’t signed, but the appetite is.

Softbank’s piggy bank rattles
SoftBank, which took Arm private in 2016 and still owns 90% of the economic value, watched its own Tokyo shares leap 7.9% on the news. The windfall gives Masayoshi Son fresh firepower for his OpenAI and Stargate data-center bets, projects that will, coincidentally, consume mountains of Arm CPUs. The snake eats its own tail—then sells the tail back at 50% gross margin.
Analysts who spent the last decade asking when Arm would grow into its valuation suddenly face the opposite problem: the stock trades at 28× the new 2029 earnings target, a multiple that assumes zero execution risk and zero competitive response. Intel, still bleeding server share, and AMD, now fighting a two-front war, will not applaud politely.
Bottom line: Arm has spent three decades renting out the blueprints; now it is moving into the house, raising the rent and changing the locks. The queue outside suggests the tenants are willing to pay. Whether the incumbents file eviction notices is the next act—and it will be expensive for whoever loses.
