ByteDance’s AI chip talks with Samsung Foundry
Reuters reports that TikTok owner ByteDance is developing an AI inference chip with Broadcom and is in talks with Samsung to manufacture it—an apparent shift from earlier expectations that the chip would be made at TSMC. Here’s a neutral, Reuters-style news analysis of the reported timeline, volumes, and strategic implications.
ByteDance, the Chinese owner of TikTok, is developing a custom chip designed primarily for artificial intelligence inference workloads and is in talks with Samsung Electronics to manufacture it, Reuters reported, citing people familiar with the matter. The report said ByteDance is collaborating with U.S. chip designer Broadcom on the project and aims to receive engineering samples by the end of March 2026.
The discussions, if they progress to production, would place Samsung Foundry—rather than Taiwan Semiconductor Manufacturing Company (TSMC)—in the manufacturing role that earlier reporting had suggested. The shift matters because it touches three pressure points in the current AI build-out: foundry capacity and ramp certainty, advanced packaging and memory availability, and the growing list of companies trying to reduce dependence on a single class of accelerators.
What we know from the reporting
According to Reuters, ByteDance’s chip is intended primarily for AI inference—the “serving” side of AI where trained models are deployed to generate responses, rank content, moderate video, and power recommendation engines. Reuters said ByteDance is in talks with Samsung for foundry manufacturing and expects to receive sample/engineering chips by late March 2026.
The report also described ByteDance’s production ambitions: the company aims to have at least 100,000 units produced in 2026 and to ramp output to around 350,000 units over time. Reuters added that the manufacturing discussions may include memory chips as part of the package—an important detail because memory, particularly high-bandwidth memory used in AI systems, has become a key constraint as AI demand rises.
ByteDance denied the chip project in a statement cited by Reuters, the report said. Still, the sourcing and specificity around timing, volumes, and supplier conversations place the story in the category of “supply-chain in motion” rather than a generic exploration.
How this connects to the earlier TSMC expectation
In June 2024, Reuters reported that ByteDance was working with Broadcom to develop an advanced AI processor, with manufacturing expected at the time to be outsourced to TSMC. That earlier report framed the effort as a way for ByteDance to secure access to high-end chips amid U.S.-China tensions and a global surge in AI compute demand.
The new reporting does not automatically mean TSMC is “out”—companies often evaluate multiple foundry paths and timelines before committing. But it does suggest that Samsung has become a serious candidate in ByteDance’s supplier calculus as the chip moves closer to samples and potential ramp.
Why inference chips are where the economics turn
AI training often dominates headlines because it involves massive clusters and frontier models. In mature products, however, inference typically becomes the enduring cost center: every user interaction that calls a model—search, recommendations, translation, content understanding—creates an ongoing stream of compute demand.
That is why large platforms increasingly consider custom silicon: if your workload is predictable and large enough, a tailored accelerator can lower cost-per-inference, reduce latency, and stabilize capacity planning. ByteDance’s core businesses—short-form video distribution, recommendation, ad targeting, and safety moderation—map naturally onto inference-heavy pipelines.
“Inference is where AI becomes a utility bill.”
— Industry shorthand for the shift from one-time training to recurring serving costWhat Samsung Foundry might offer
Reuters’ reporting foregrounds two elements: (1) Samsung as manufacturing partner, and (2) the possibility that memory supply is part of the broader commercial package. Together, these imply ByteDance is not only thinking about a compute die, but about a deliverable system component that must land in data centers on a schedule.
The memory detail is not trivial. High-bandwidth memory (HBM) is critical to modern AI acceleration, and Reuters has separately reported on the intensity of competition in next-generation HBM as demand for AI data centers expands. Samsung itself has publicly emphasized strong memory demand extending into 2027, and Reuters reported Samsung has shipped newer generations of HBM as the company tries to catch up in the market.
Analysis: Why a “Samsung path” could be attractive
If ByteDance is indeed leaning toward Samsung Foundry, there are a few plausible incentives consistent with industry behavior—none of which are confirmed in the Reuters report, but all of which fit the current AI supply-chain reality:
- Ramp certainty and scheduling leverage: For advanced silicon, the timeline from samples to volume production depends on foundry allocation, yield learning, and packaging readiness. A second credible foundry option can improve negotiating leverage on timelines and commercial terms.
- Bundled negotiation across compute and memory: If memory chips are part of the talks, ByteDance could be trying to reduce the risk that it gets compute dies without a matching memory supply window.
- Strategic diversification: AI infrastructure has become sensitive to supplier concentration risk. Even if a custom chip does not replace GPUs, it can reduce exposure to shortages and policy-driven constraints by shifting a portion of serving workloads to internal silicon.
The key point is not that Samsung is “better” than TSMC in the abstract. It’s that, for a specific product at a specific moment, the winning supplier can be the one that can credibly commit to a package: node + yield curve + packaging + memory alignment + delivery schedule.
How big is “100,000 chips,” really?
In AI, raw unit counts can be misleading without context on die size, packaging complexity, and deployment model. But even as a headline figure, 100,000 units in a year is a non-trivial target for a custom accelerator—implying real internal confidence in workload fit and platform readiness.
The eventual ramp figure—around 350,000 units—would, if achieved, move the chip from a niche experiment toward a material part of ByteDance’s AI serving footprint. That said, the step from “plan” to “shipped, deployed, and cost-effective” is where most custom silicon initiatives face their hardest tests: software enablement, compiler maturity, operator tooling, and integration into existing inference stacks.
Analysis: What this could mean for Nvidia’s dominance
The most likely outcome is not a dramatic displacement of Nvidia GPUs, but a gradual segmentation: GPUs remain essential for training and for rapidly evolving model architectures, while custom inference chips absorb stable, high-volume serving workloads where cost and latency dominate.
This pattern already exists elsewhere in the industry: companies build or buy inference accelerators to reduce cost-per-request once product usage becomes predictable at scale. If ByteDance can move even a fraction of its inference to custom silicon, it could lower unit cost and smooth capacity planning, while continuing to rely on GPUs for frontier training and experimentation.
What we still don’t know
The reporting leaves several technical and operational details undisclosed—details that will determine whether the chip becomes a major platform element or remains a limited deployment:
- Process node and performance target: No confirmed node, performance-per-watt target, or pricing model has been reported in this update.
- Packaging approach: Whether the chip uses advanced packaging solutions and what memory interface is planned has not been detailed.
- Software stack readiness: Compiler toolchains, kernel libraries, and integration into existing inference services typically decide real-world success.
- Deployment scope: Which internal services (recommendation, moderation, ads, search, multimodal pipelines) would be prioritized is unknown.
- Supplier roles: Whether Samsung would be sole foundry partner, second-source option, or initial ramp partner is unclear.
What to watch next
If Reuters’ timeline is directionally correct, the next near-term checkpoint is late March 2026 for engineering samples. After that, the signals that matter most will be operational rather than promotional:
- Sample-to-qualification cadence: How quickly the project moves from engineering samples to qualification lots will indicate seriousness and maturity.
- Packaging and memory alignment: Any evidence that memory supply is secured for the same deployment windows as compute silicon will reduce execution risk.
- Volume commitments: Whether “100,000 units in 2026” becomes purchase orders, production starts, and shipments.
- Internal platform messaging: In large companies, the language shifts from “experiment” to “platform” when a chip is integrated into standard deployment templates.
Bottom line
Reuters’ reporting suggests ByteDance is pushing deeper into the hyperscaler playbook: build internal silicon for inference, reduce supplier concentration risk, and potentially secure memory in parallel. The noteworthy twist is the reported foundry conversation with Samsung—an apparent departure from the earlier expectation of a TSMC manufacturing path.
If the project reaches production at the scale described, it would underscore a broader AI reality: the next phase is not only about model quality. It is also about industrial execution—supply chains, packaging, memory, cost-per-inference, and the ability to deploy AI as a durable, high-volume service.
Sources
- Reuters (Feb 2026): Report on ByteDance developing an AI inference chip with Broadcom and talks with Samsung for manufacturing; samples expected by end of March; production targets cited.
- Reuters (Jun 2024): Earlier report on ByteDance and Broadcom working on an advanced AI processor and expectation of TSMC manufacturing at the time.
- Reuters (Feb 2026): Coverage highlighting ongoing strength in AI-driven memory demand and competition in next-generation HBM shipments.
This post summarizes public reporting and provides clearly-labeled analysis. It is not investment advice.