11/16 2026
384
Preface
After the market close on November 4, Eastern Time, Qualcomm released its Q4 fiscal 2026 results. Handset chip revenue came in at $5.137 billion, down 26% year-over-year, and profit guidance for the next quarter fell short of Wall Street expectations. Following the earnings release, the stock plunged nearly 10% at one point.
Forty-eight hours later, the plot reversed. On November 6, a 13F filing submitted to the SEC by Google's parent company Alphabet revealed that, as of September 30, the search giant held approximately 2.977 million Qualcomm shares, worth about $548 million based on the quarter-end stock price. It was a first-time position, accounting for less than 0.5% of Qualcomm's total share capital. Once the news broke, Qualcomm rose about 5% in premarket trading.
A company being pulled back and forth by two documents within three days—this image itself is a metaphor for today's chip industry. The old engine is still leaking oil, yet the ignition of the new engine is already clearly audible.
Author | Fang Wensan
Image Source | Internet

What weight does $548 million carry on Alphabet's books?
First, let's clarify the coordinates of this position. Alphabet's total holdings market value in Q3 was approximately $88.3 billion, down 10.9% quarter-over-quarter. SpaceX alone accounted for about $83.734 billion, or 94.78%, and the top ten holdings together made up 99.48%.
In such a highly concentrated portfolio, Qualcomm ranked as the sixth-largest holding with a weight of about 0.62%, right next to a similarly sized Arm position of about $568 million.
This position itself carries directional significance. During the quarter, Alphabet added 4 stocks, increased 1, reduced 3, and fully exited 3 (GitLab, Spero Therapeutics, and Autolus Therapeutics). The overall portfolio was shrinking, yet Qualcomm entered the top six directly as a new target.
A 0.62% position is so light relative to Alphabet's balance sheet that it is almost negligible—too light to be explained by ordinary asset allocation logic. Holding both an IP company and an SoC company at the same time sketches out a map pointing to the ecosystem foundation upstream in chips and terminal compute.
There is another detail worth noting: the filing shows this Qualcomm position is managed by GV Management Company, Google's venture capital arm. GV has long operated relatively independently, with a style somewhere between financial return and industrial positioning.
Reading this trade as a strategic declaration from Google headquarters may be overreach, but reading it merely as a financial allocation is clearly insufficient.
The factual boundary that must be drawn is this: a 13F is a static snapshot delayed by 45 days. It only shows what is held, not why it is held. Equating establishing a position directly with placing an order is like treating an option as a contract.

The skeleton is standing, but the flesh and blood are still on the way
Before Google's capital entered, Qualcomm's data center map had already completed a round of intensive moves.
At the New York investor day in June this year, the Dragonfly product portfolio was unveiled in full. The data center CPU C1000 uses self-developed Oryon cores, scaled to more than 250 cores with clock speeds above 5 GHz, paired with PCIe Gen 7 and CXL. The inference accelerators AI200, AI250, and AI300 iterate on an annual cadence, alongside connectivity chips and custom silicon businesses.
Financial targets followed: $5 billion in data center revenue in fiscal 2027, $15 billion in fiscal 2029, total non-handset revenue of $40 billion in 2029, and adjusted earnings per share above $18.
The quality of the customer list exceeded most people's expectations. Meta confirmed adoption of the C1000, with deliveries starting in the second half of 2028; Microsoft signed on for a new category of AI chips based on HBC technology; Azure has confirmed deployment.
The agreement with Amazon AWS in September went further, advancing custom chips, AI inference, and 1.6T optical interconnect as a bundle. Qualcomm simultaneously issued warrants to Amazon at an exercise price of $161.26, corresponding to up to 25 million shares, or about $4 billion. An initial 3.75 million shares were granted first, with the remainder unlocked with commercial milestones, tied precisely to procurement itself, with a ten-year cumulative cap of $60 billion.
Revenue recognition has already begun. Data center revenue in the fourth fiscal quarter was about $650 million, above market expectations, mainly from custom and connectivity chips. The company expects continued sequential ramp-up in the current quarter, a full-year fiscal 2027 target of $5 billion, and an annualized run rate potentially reaching $7 billion by fiscal year-end.
The other side of these numbers is the accelerating blood loss in the handset base. Rising memory prices have pushed up full-device bill of materials costs, suppressing shipments of mid- and low-end models, and QCT pre-tax margin fell from 29% in the year-ago period to 24%.
Apple's replacement with its own modem is entering the sprint phase. Revenue from Apple in the first quarter of fiscal 2027 is expected to halve quarter-over-quarter and is no longer included in guidance.
The automotive business posted $1.990 billion in a single quarter, up 89% year-over-year, with an annualized run rate of about $8 billion—another pillar that has timely stepped up.
Here lies a rather ironic loop: AI data centers have swallowed memory capacity, rising memory prices have in turn wounded Qualcomm's handset business, and one of the remedies Qualcomm offers happens to be reinventing the memory architecture.
The speed at which the old business is losing blood and the speed at which the new business is transfusing it will form the main melody of Qualcomm's valuation over the next two years.
The stock has risen more than 45% over the past six months, indicating that the market has already prepaid part of the deposit for the transformation. The nearly 10% selloff after earnings also shows that this deposit may be called for a top-up at any time.

Two grammars of alliance
Looking at Amazon's and Google's deals together reveals two grammars by which cloud giants bind chipmakers.
Amazon's warrant is tight: exercise price, milestones, and a ten-year $60 billion cap write procurement into a bet, essentially using equity to lock the customer onto its own roadmap. The more it purchases, the more it unlocks, welding both parties' interests together.
Google's stake is loose: no exercise price, no procurement terms, not even a public statement. One is like a marriage contract; the other is like exchanging business cards for the first time.
Between tight and loose, what is reflected is the transformation of the entire industry's relationship structure. Ten years ago, there were only procurement contracts between cloud vendors and chipmakers—whoever's product was better was the one bought.
Then came the self-development wave: Google TPU, Amazon Trainium, Microsoft Maia, pulling the core compute back into their own hands.
What is happening now is the third step: investment, customization, and multi-source procurement combined into a combo punch. Self-development cannot fill the demand gap of the inference era, so cloud vendors are beginning to use capital ties to reserve seats on chip companies' roadmaps.
Google's compute portfolio is already spread wide: self-developed TPU is one line, a cooperation agreement with Marvell worth up to $12.2 billion is another line, holding Arm shares is another line, and now Qualcomm adds one more square.
The four lines point to the same demand: no matter which supply chain has problems, the supply of inference compute cannot be interrupted. The timing is also intriguing. Google's position was established after Microsoft signed HBC and Amazon's warrant was finalized; it fills the square it had been absent from.
Thus a judgment can be made about the nature of this investment: what Google bought is not Qualcomm's present. A company whose handset chip revenue fell 26% cannot deliver a quarterly report that excites Google.
This $548 million is closer to a strategic option, whose underlying is whether Qualcomm can become the second source of compute for inference infrastructure.
The exercise method is unknown; it could be a future procurement contract, or it could remain a financial holding forever. The option's validity depends on HBC.

HBC: moving the stove into the granary
In the large-model inference stage, every time a token is generated, the chip must move massive model parameters along with an ever-expanding key-value cache from memory into the compute unit. The arithmetic itself is cheap; the moving is expensive.
Industry estimates cited by Qualcomm show that over a decade, transformer model parameter scale has expanded about 240 times every two years, while AI hardware memory capacity has grown only about 2 times. The chef's hand speed has doubled again and again, but the width of the serving channel has barely changed. This is the physical reality of the memory wall.
HBM is the mainstream answer of this era: stack DRAM dies vertically and package them together with the GPU using a silicon interposer. The channel does become wider, but the cost is clearly priced.
HBC takes a different approach: rather than widening the serving channel, it moves the stove into the granary.
Multilayer LPDDR dies are stacked vertically through through-silicon vias, with a base die capable of actual computation pressed at the very bottom. A near-memory acceleration unit absorbs the operations with the greatest bandwidth pressure inside the stack and passes only the results to the main SoC.
A Qualcomm engineer's analogy is very direct: think of it as a coprocessor. What travels over the link need not be raw data; it can be computed results.
The data movement distance is compressed from centimeters to micrometers, and the entire solution does not require a silicon interposer. A standard organic substrate can carry it, bypassing CoWoS, one of the most capacity-constrained links globally.

At the numerical level, first-generation HBC brings the AI250 768 GB of memory per card and 133 TB/s of effective memory bandwidth, about 18 times that of the AI200 using LPDDR5X.
Bandwidth per watt exceeds the expected HBM4 specification by 5 times, with the company's stated ultimate goal at 6 times. Storage capacity per unit of power is about 200 times that of on-die SRAM.
In November, the San Diego lab announced that first-generation HBC completed hardware validation. The full read-write path and near-memory computing capability were confirmed at production-level dimensions, and the testing still used industry-standard ATE equipment, making the validation process itself reviewable.
Beneath the parameters, HBC's underlying logic is an asset equation. Qualcomm's foundation for doing this is the muscle memory built over 30 years in the handset business.
Low-power design, the LPDDR supply chain, packaging experience in TSV and wafer-level bonding, and architectural accumulation in Hexagon NPU are all legacy assets from the mobile era.
Using the mobile ecosystem's cheap, high-capacity memory to attack the data center ecosystem's expensive, scarce memory is a textbook flanking maneuver.
Nvidia's moat lies in peak compute and CUDA software gravity. Qualcomm bypasses the frontal battlefield and focuses specifically on the cost of data movement.
In the inference era, the bulk of the cost per token is determined by memory. If this route works, what gets rewritten is not only Qualcomm's own valuation but also the bill of materials of an entire inference data center.
Qualcomm has also publicly disclosed a farther-reaching intention: to bring this stacked architecture back to Snapdragon so that on-device AI can enjoy the same energy-efficiency dividend. That would be an application surface broader than data centers.
Of course, restraint is necessary. HBC trades peak bandwidth density for capacity and energy efficiency, and the training side will remain HBM territory in the near term.
A more likely picture is layered coexistence: training relies on HBM, while on the inference side HBM, LPDDR stacking, and near-memory computing race side by side.

Conclusion
Four things are worth watching next: whether the AI250 can sample on schedule and deliver on its promised bandwidth per watt under real workloads; whether Google increases, holds, or exits in the next 13F; and the harder signal—whether Google will become a named customer of Dragonfly.
$548 million cannot buy a strategic high ground, but it is enough to buy the observation seat with the best view. Google has already taken its seat, and the real show is set for 2027.
Some materials referenced: U.S. Securities and Exchange Commission (SEC): "Alphabet Inc. Q3 2026 Form 13F Holdings Report"; Qualcomm: "Qualcomm Announces Fourth Quarter and Fiscal 2026 Results"; Alphabet: "Alphabet Announces Third Quarter 2026 Results"; Reuters: "Alphabet takes $548 million stake in partner Qualcomm in data center push"; Futurum Group: "Qualcomm Validates First HBC Silicon With an 18x Memory Bandwidth Gain Over LPDDR5X"