AI's Next $1.3 Trillion Bet: 5 Stocks Poised for the Inference Wave

Bloomberg Intelligence's $1.3 Trillion AI Picture Points Toward Inference

This is the shift investors need to anchor on: the AI buildout began as a training story, but the longer-term commercial prize looks more like inference. Bloomberg Intelligence sees the generative AI market growing to $1.3 trillion over the next 10 years, with growth initially driven by training infrastructure and later shifting toward inference devices for large language models. Training is the expensive classroom; inference is the daily usage that keeps running after the model is deployed.

Why timing matters more than the headline number

The market is still pricing a nearly endless spending boom, but the first signs of a slowdown are appearing. The five largest US cloud and AI infrastructure providers are guiding to roughly $660 billion and $690 billion on capital expenditure in 2026, and other company-level guidance adds up to about $725 billion in 2026. If spending gradually leans more toward inference-heavy products and services, the companies that capture ongoing compute usage could be well placed.

The bear case matters too. UBS expects hyperscaler spending growth to slow sharply over time, to just 6% in 2028. That makes the inference turn not just a technical debate, but a real investment crossroads.

Inference changes the revenue model: from one-off builds to repeated usage

A market that is already sizable and still growing

The inference market was USD 103.73 billion in 2025 and is projected to reach USD 312.64 billion by 2034. That matters because inference is not a single stadium you build and open. It runs through clouds, networks, devices, apps, and APIs. For investors, that can make revenue broader and more recurring, and less dependent on one-off capital projects.

Open-weight models widen the hardware field

When Moonshot AI released Kimi K3 as an open-weight model, part of the sell-off reflected a simple fear: if models become cheaper and easier to run locally, maybe the hardware buildout cools. The more nuanced read is that inference may be less centralized than training. That does not kill the hardware opportunity; it spreads it across more architectures and deployment models.

Nvidia still has a commanding position in AI acceleration, and Blackwell offers 25 times better energy efficiency. But inference also spans edge, cloud, and on-premises systems, which is why the competitive field is broader than the training story alone would suggest.

Five stocks positioned for a broader inference market

If the money pool shifts from model training to ongoing inference, the most direct exposure sits in accelerators, networking, and manufacturing. The five names below sit across those layers.

Nvidia: still the core inference accelerator

Nvidia remains the first name because its GPUs already power large language models, and Blackwell offers 25 times better energy efficiency. Bank of America also still views Nvidia as a core "AI compute" leader.

What could break it: UBS expects hyperscalers' capex growth slowing to 6% in 2028. If spending cools and hyperscalers rely more on custom silicon, Nvidia can still win on performance, but its share and margins may become harder to defend.

AMD: a more flexible pick if inference fragments

AMD benefits if inference demand spreads across edge, cloud, and on-premises systems. Bank of America also lists AMD among its "AI compute" leaders, which supports the case that it can participate if customers want capable inference options at different price points and locations.

What could break it: A faster slowdown in hyperscaler spending, or a stronger customer shift toward Nvidia or proprietary chips, could pressure AMD's mix.

Broadcom: the networking layer tied to real-time demand

Bank of America says Nvidia and Broadcom continue to power those AI ambitions, with gains driven by compute, networking, memory. As inference becomes a continuous stream of requests rather than a one-time buildout, the network layer can become increasingly important.

What could break it: If capex cools enough that customers delay network upgrades or simplify architectures, Broadcom's growth path could slow ahead of expectations.

Marvell: a more tactical way to play widening inference spend

Bank of America is also doubling down on Marvell Technology (MRVL) as one of its "AI compute" leaders. That makes Marvell a plausible way to express the idea that inference spending may broaden beyond the headline accelerator market.

What could break it: If platforms keep more design, procurement, and margin in-house, Marvell may not capture as much of the usage ramp investors expect.

TSMC: advanced manufacturing remains a bottleneck

TSMC is stepping up production of cutting-edge 3-nanometer and 5nm chips to meet surging demand. Even if the inference winner set broadens, advanced foundry capacity still acts like a toll booth on the stack.

What could break it: Faster in-house silicon adoption, slower adoption of leadership process nodes, or foundry execution issues could all weaken that choke-point logic.

A practical way to think about exposure is: Nvidia and TSMC look like the cleaner core positions, while AMD, Broadcom, and Marvell may offer more upside but with more volatility.

How to approach the trade if capex normalizes

UBS expects hyperscaler capex growth to fall from 76% this year to 25% in 2027 and then 6% in 2028. At the same time, 82% viewed semiconductors as the market's most crowded trade. That argues for discipline: the setup is not to buy anything with a chip attached, but to wait for expectations and economics to reset as the spending ramp normalizes.

What would support the thesis, and what would challenge it

• Support: The cash being deployed is still enormous, with about $725 billion in 2026 guided across major platforms.
• Support: Investors are already scrutinizing that spend more closely, with shares of Amazon, Meta, and Microsoft falling after Alphabet hiked its 2026 capex forecast.
• Challenge: A hard break in AI investment. Reuters' analysis still points to $340 billion more in annual operating cash flow in 2027 than in 2025, which argues for deceleration rather than a collapse.
• Challenge: A fade in inference demand. The market was USD 103.73 billion in 2025 and still tracks toward USD 312.64 billion by 2034.

The practical stance is to use the next earnings cycle and the 2027 capex debate as the decision window, not the next headline surge.