After 145X on AMD: Three Stocks Running the Same Algorithm — AbCellera, AST SpaceMobile and Rocket Lab

Welcome back to the world's best investing podcast — Antonio Linares opens there, reminding viewers that his AMD position started in August 2014 is now up 145 times the original outlay. The thesis has been public since around 2020, written up in detail in 2022 when the consensus feared a CPU squeeze, and it aged well: in November 2021 he called for a trillion-dollar market value within five years, AMD crossed 1.01 trillion yesterday, two months ahead of schedule. His five-year target of 600 per share, first floated for 2023, was also cleared early. For Linares the lesson is directional accuracy, not price precision, and he flags the whole showcase as educational, not financial advice.

Why Wall Street misread AMD: machine, not product

The mistake, in his telling, was benchmarking AMD's latest product against Nvidia's latest product. That narrow comparison kept investors selling AMD at 75 last year while he kept publishing that the company was becoming a trillion-dollar AI platform. What worked instead was understanding chiplet architecture: compute is distributed by nature, so a mix-and-match engine that assembles compute tiles on demand is the cheapest way to lower the cost per token as scaling continues. AMD's bet on Infinity Fabric, fast SerDes and advanced packaging is the physical expression of that idea, and it is hard for a monolithic die to follow once tokens become the unit of account for the economy.

Linares calls the pattern the AMD algorithm and stretches it beyond semiconductors, with Tesla as a parallel: start from a hard, expensive, fragmented industrial problem with potentially endless demand on the other side, build a modular, vertically integrated machine around it, accumulate proprietary process knowledge, add software and data feedback loops so the machine improves itself, then manufacture at scale. Over time the economic composite grows — more variables, more use cases, adjacent markets no one forecasted — while marginal cost falls and the machine throws off rising free cash flow per share. Stock prices, he notes, follow that per-share cash flow in the long run. That is why he sees AMD as just beginning, with a path to five or even ten trillion over the next decade as custom silicon is stamped at ever lower incremental cost.

AbCellera: Lego logic in biology

The same template in biology is AbCellera. At the protein level the body looks like Lego; if you can find the right push buttons with precision, you can in principle do a great deal that was practically impossible until recently. Every acquired capability on the platform — each engineered module added through deals — can be reused many times across many future drug programs, spreading the fixed cost of discovery over an expanding surface. Linares frames it as the Palantir of biotech in that sense: value compounds in process, not in any single asset seen today.

The clearest commercial proof in Q2 2026 came from the T-cell engager platform. Linares cites an 84 million upfront cash payment from Jazz in the quarter; the formal news release dated June 17, 2026 describes 56 million upfront and up to 792 million in option and milestone payments for any program Jazz elects to take forward, with management talking about more than 2 billion in potential downstream payments and up to about 4 billion in total deal value. The numbers differ in presentation but point the same way: the platform drew real cash for the first time at scale, helped lift cash from operations in Q2, and marks the first quarter where the discovery engine looks like a commercial platform rather than a science project. AbCellera remains early in financial terms, but the pipeline now funds itself in part.

Where the pipeline stands

Risk is still binary on the candidates. ABCL575 completed its single-ascending-dose Phase 1 and awaits readouts, ABCL615 — a selective NK3R modulator for vasomotor symptoms such as hot flashes — is approaching a data window, and ABCL688 and ABCL386 are in Phase 1 and Phase 2 respectively. The TCE biology recap is quick: there are two ways to get T cells to kill. One, illustrated by ImmunityBio, is to teach the immune system via MHC-mediated peptide presentation after IL-15 super-agonist reactivation; the other, AbCellera's route, is to engineer any T cell to bind a chosen target directly. Linares sees the engineering route as more directly steerable, and the size of the upfront suggests the market is starting to price that distinction. He still stresses that candidates are unproven and the platform, not any single readout, is the bet.

AST SpaceMobile: not a satellite company but a spectrum company

AST SpaceMobile looks like a satellite company and trades that way, but the thesis Linares pushes is that it makes custom spectrum. As billions more devices — phones, IoT sensors, defense nodes and objects in orbit — come online, spectrum stops being a fixed commons and has to be dynamically assigned per device, location and workload, because each device has distinct physical constraints for delivering value. The answer is direct-to-device with huge phased-array satellites that can broadcast any spectrum to billions of devices at once; each satellite is effectively a giant cell with immense capacity that extends existing terrestrial networks rather than replacing them.

Q2 2026 numbers suggest that bridge is being built. Revenue was 31.5 million and rising, with 13 gateways delivered to seven customers across five continents and work underway on 50 gateways in 20 markets. Production is running at six satellites per month and the target is 45 satellites in orbit by early 2027, at which point the constellation integrates with carrier networks so a handset needs only a text-message consent to roam onto space. In the United States more than 3000 low-band cells have been lit as logical service areas that let ASTS satellites behave as an extension of the ground grid, with a path to 5600. Those cells are not towers but digital infrastructure; the newest ASIC brings ten times the compute of the prior engine, and management has for the first time started talking about later verticals — radar, emergency response, IoT, sovereign infrastructure and AI edge compute that could eventually involve AMD's FPGAs.

Rocket Lab: from launch service to infrastructure

Rocket Lab's leg of the story has always been Neutron turning launch from a service into infrastructure. The small Electron vehicle proved the team can fly frequently; the medium-lift Neutron is meant to change unit economics by putting more mass in orbit per flight. Its Archimedes engine — 3D-printed, methane and liquid oxygen, oxidizer-rich staged combustion, designed for up to 20 reuses — has now logged more than 400 hot-fire tests at NASA Stennis, and every major vehicle element is in advanced test or trial assembly. Guidance stays at a first Neutron launch in Q4 2026, with management stressing they are optimizing for flight ten, not just flight one, the same marginal-cost logic that defines the AMD algorithm. Customers are booking Neutron before it flies, including the U.S. Space Force for dedicated missions, at an average selling price of 50 to 55 million per rocket.

The second layer is Photon Lite, a standardized spacecraft bus — essentially a container that can turn anything that fits inside Neutron into an on-orbit asset. Peter Beck put it bluntly on the Q2 call: most of what Rocket Lab expects to do going forward will be built on that platform. The first Photon Lite booking came inside the Space Force package just mentioned, so the first integrated Neutron plus Photon contracts are already on the books. In Linares's reading that is the bridge from one-off launch revenue to durable cash: each launch becomes an enduring cash-flow source because the company can build, launch and operate constellations and then clip fees.

The spectrum race where paths cross

What ties ASTS and Rocket Lab together right now is the scarcity of spectrum. Billions of devices are already deployed and tens of billions more are coming on Earth and in orbit, and every one of them needs a different slice of spectrum delivered with different physical trade-offs. Regulatory coordination makes rights hard to replicate. The two companies attack the same scarce pool from opposite sides: Rocket Lab builds launch and spacecraft first and then adds spectrum, ASTS builds a satellite proof that now wants launch, compute and applications. The cost per satellite at ASTS even spiked for a while because the company paid for priority access to the launch manifest. Linares's point is that we have seen this convergence before across semiconductors and Web 2.0: many players chase one valuable stack, each masters a different entry wedge enough to stay alive, generates free cash that funds the next wedge, and in a decade no one can cleanly separate the winners by today's product label. The bottom line stays that the visible product is not where compounding lives — AMD, AbCellera, ASTS and Rocket Lab each compound in the machine behind the product — and markets that price only the last quarter miss that shift, as AMD at 70 last year showed. AbCellera has moved on Jazz cash, ASTS and Rocket Lab remain roughly 50 percent discounted with underlying value doubling since the last selloff, yet early financials mean Neutron, cadence and trial readouts can all slip.