AWS taps Qualcomm for custom inference silicon and 1.6 Tbps optics
On September 8, 2026, Amazon Web Services announced a partnership with Qualcomm for custom AI inference silicon and 1.6 Tbps optical interconnect gear, under a commercial commitment that could reach $60 billion through September 2036. A strong signal about the diversification of AWS’s silicon supply chain — and about the real bottleneck of AI clusters: the network.
September 8, 2026. Amazon Web Services announced it has tapped Qualcomm to develop “customized” silicon for AI inference and 1.6 Tbps optical networking gear, while offering few technical details. The commercial commitment — purchases and other consideration — “could total up to $60 billion” through September 2036. Why it matters: AWS, which already fields its own Graviton, Trainium and Inferentia chips, is widening its supplier base again, and the choice of Qualcomm says as much about networking as about compute.
A fourth silicon supplier joins AWS’s game
AWS is no stranger to custom silicon. It has already had customized variants of Intel Xeon processors built for it, and AMD has produced custom CPUs such as the Epyc 9V64H for Microsoft. Add the in-house chips — Graviton for general compute, Trainium for training and Inferentia for inference — and the arrival of Qualcomm draws a four-source portfolio, a level of supply diversity no other hyperscaler currently matches.
Qualcomm’s appeal is twofold. First, its emerging datacenter lineup: the AI250 “Dragonfly” platform, unveiled at its investor day earlier this summer, swaps pricey HBM for what the company calls “high-bandwidth compute” — cheap LPDDR memory with some compute shifted onto the memory’s base die. Qualcomm claims a higher “effective” memory bandwidth at lower cost, aimed precisely at the inference bottleneck. Second, a future datacenter CPU, the C1000, announced at 5 GHz and more than 250 cores for the second half of 2028. It is plausible that AWS receives a customized AI250 variant, just as it did with Intel for Xeon 6.
The real signal: 1.6 Tbps optics
The compute side of the partnership remains deliberately vague — The Register noted that the canned statement from Qualcomm CEO Cristiano Amon says nothing about the deal’s nature. The networking side is unambiguous. AWS will use Qualcomm’s SerDes and optical DSPs, inherited from the late-2025 acquisition of Alphawave Semi, to push port speeds to 1.6 Tbps.
That is where the strategic story lives. As an AI cluster scales from a few dozen accelerators in a rack to hundreds or thousands spread across racks, optical interconnect becomes the dominant bottleneck — not compute. Optical pluggables and co-packaged optics are now a major supply-chain pressure point for scaling AI infrastructure. Adding Qualcomm to its optical component sources lets AWS secure a critical link for its next generations of clusters.
The relationship runs both ways: Qualcomm plans to expand its use of AWS’s Bedrock platform for electronic design automation (EDA) workloads. It is the classic loop of AI deals — the silicon customer also becomes a cloud customer — but it signals that Qualcomm is betting on AWS models to speed its own design flows.
What it means for cloud buyers
The announcement has no immediate impact on pricing or the instance catalog: AI250 silicon is expected only next year, the C1000 in 2028, and no product has been named. But it lights up three trajectories.
On inference. A fourth chip supplier in AWS’s fleet strengthens an already-running trend: a multiplication of low-cost inference instance options. For stable, predictable inference workloads, competition between Graviton, Trainium/Inferentia, NVIDIA and soon Qualcomm translates into downward price pressure — provided you accept the work of benchmarking and porting onto the new families.
On cluster networking. The move to 1.6 Tbps interconnects sets a direction for high-performance-computing and distributed-training services. Teams designing multi-rack clusters should fold this optics generation into their capacity plans now, or risk discovering the bottleneck is no longer the GPU but the cable.
On supply sovereignty. Diversifying beyond Intel, AMD and NVIDIA reduces dependence on any single supplier of critical components. That is an engineering decision, but also a continuity one: a hyperscaler controlling four silicon sources is less exposed to a shortage or a sectoral embargo.
A silicon strategy that did not start yesterday
The Qualcomm deal fits a long-running line. AWS launched Graviton in 2018, an in-house Arm chip for general compute, now in its Graviton5 generation; Inferentia for inference in 2019; Trainium for training in 2021, scaled up through the Trainium2 and Trainium3 generations. Add the Nitro System, the hardware hypervisor that virtualizes instance networking, storage and security, and the strategy has steadily reduced AWS’s dependence on NVIDIA for the workloads it controls end to end. In-house compute now covers a growing slice of the instance catalog.
Qualcomm’s pitch reads well in that context. The AI250’s high-bandwidth compute swaps HBM — expensive, scarce and dominated by a few suppliers — for cheap LPDDR, shifting part of the compute onto the memory’s base die. The trade-off is explicit: LPDDR delivers less raw bandwidth than HBM, but a far lower cost per gigabyte. For inference, where the binding constraint is often cost per token rather than peak training throughput, that trade-off can win.
The choice also reveals a market truth. The cost of HBM has exploded with accelerator demand, to the point of becoming a line item of its own in a cluster’s bill. Any credible alternative that loosens that dependence, even at the price of lower bandwidth, mechanically interests hyperscalers. Qualcomm is not alone on this front either: the race to alternative memory architectures is one of the most contested corners of AI silicon. Whether the promise holds on real workloads is a question independent benchmarks will settle when the first instances ship.
Why the optics matter more than the chips
The optics detail deserves its own weight. Co-packaged optics — moving the optical engine onto the switch ASIC’s package — and near-packaged optics are where the industry is heading to escape the power and latency cost of pluggable modules at 1.6 Tbps. Qualcomm’s SerDes and DSP portfolio, built on the Alphawave Semi acquisition, positions it as one of the few merchant suppliers able to serve that transition. For AWS, that is a hedge on a component class that has become genuinely hard to source at scale.
This is not a luxury bet. Optical interconnects are the quiet constraint of the AI buildout: as clusters grow past a few thousand accelerators, the energy and cost of shuttling data between racks starts to rival the energy of computing itself. A hyperscaler that locks in 1.6 Tbps components early buys capacity headroom and a second source of supply, both of which translate into faster cluster delivery and lower per-rack cost.
The circularity is worth noting too. Qualcomm will expand its use of Bedrock for EDA workloads, so the chip vendor is at once a silicon supplier and an AWS cloud customer. Such loops are increasingly the norm in AI infrastructure, and they blur the line between partner and competitor in ways that matter for negotiation, and for anyone trying to read AWS’s true cost structure. None of it changes the catalog today; all of it reshapes the road map tomorrow.
Verdict
If you run inference workloads on AWS, change nothing today, but watch for the AI250 family next year: if the high-bandwidth compute promise holds, it could reset the price/performance of inference and deserve a serious re-benchmark. If you size multi-rack AI clusters, fold 1.6 Tbps interconnects into your networking roadmap now — that, not silicon, is where the next scaling limit will bite. And if you read the deal as a market signal, note that the cloud-chip war no longer pits NVIDIA against everyone else: it is becoming a four-supplier contest, where supply-chain diversification is a competitive advantage in its own right.