2026 Debates and Positioning

The central question for 2026 is whether the AI investment cycle has reached its peak or if we are still in the early innings of a multi-year build-out. Compared with prior booms or bubbles, the investment community retains a surprising amount of skepticism toward the durability and usefulness of AI infrastructure. At ACM, we believe the build-out is far from over. AI infrastructure spending continues to expand even as the narrative shifts from unlimited growth to more disciplined capital allocation. Hyperscaler disclosures and commentary at the Consumer Electronics Show in January point to AI server growth of at least forty to fifty percent this year. Unlike the speculative nature of past technology booms, hyperscalers are monetizing AI services at scale, and enterprise adoption is accelerating. Supply chain bottlenecks in advanced packaging, high-bandwidth memory, and power distribution remain constraints, which reinforces the depth of demand rather than signaling weakness. AI has not peaked. It is broadening and creating multiple investible angles as different technologies in turn become bottlenecks to growth.

Hyperscaler capital expenditures provide a useful lens into this dynamic. From 2023 to 2025, aggregate AI-related capex among the top platforms grew to five hundred billion dollars, and forecasts for 2026 suggest 30-50% growth. While some investors fear saturation, the hyperscalers themselves say they are short of necessary compute capacity and are still building foundational infrastructure for AI workloads. These investments are not discretionary; they are essential to maintaining competitive advantage in cloud and AI services and to defending share against both sovereign programs and enterprise-owned deployments.

The market questions whether incremental investment dollars will flow to hyperscale datacenters or pivot toward edge AI in PCs and consumer devices. We expect datacenter deployments to remain the dominant growth engine. AI technology continues to follow the scaling laws, with model performance improving predictably given exponential increases in computing power. With the end of Moore’s Law, however, single chip or device performance no longer sees exponential gains in performance. Advanced AI requires capability that can only be found in large and growing clusters of servers in datacenters, and the gap to what can be done on a consumer device is widening. Edge devices will continue to be only a front end to the cloud for the most useful models. Edge AI will make headlines with the PC and smartphone industries touting AI enabled devices, but lack of incremental functionality will limit consumer adoption. Bill-of-materials inflation and margin compression also limit near-term profitability for device makers. Edge AI is fundamentally a volume story rather than a margin story. Our positioning remains overweight datacenter infrastructure, with edge AI viewed as optionality and a longer-dated monetization curve rather than a near-term profit pool.

The market is also debating whether the next leg of returns will come from infrastructure providers or users of AI. We remain disciplined. Infrastructure offers tangible earnings visibility, while many supposed ‘AI beneficiaries’ lack clear monetization paths. We avoid style drift into software or internet names and continue to focus on complexity in hard tech as a source of alpha.

Within AI infrastructure, the hottest debate is where to allocate capital across GPUs, ASICs, memory, and networking. We believe GPUs will maintain dominance at least through Nvidia’s Rubin platform. Custom ASICs will grow, but economics favor GPUs for general-purpose workloads that require flexibility. Compute and networking remain the anchor of our portfolio, but we see outsized opportunities in supporting technologies. Memory continues to be an underappreciated theme for 2026. The high-bandwidth memory super-cycle is real, with prices increasing at pace not seen since the mid-90s, while hyperscaler long-term agreements lock in pricing power. The memory industry needs to expand to meet multi-year demand forecasts, and this is driving a new surge in spending on capital equipment. Capital equipment companies see strong tailwinds from capacity expansion along with continued intensive growth from three-dimensional transistor structures and advanced packaging. Networking is also a battleground. Ethernet scale-up through UALink versus InfiniBand and NVLink is a structural debate, and merchant silicon vendors together with optical component suppliers offer asymmetric upside as bandwidth becomes a choke point. Our barbell approach emphasizes memory suppliers and networking plays alongside core GPU exposure.

Networking deserves special attention because it is under-owned and increasingly critical. We see 2026 as a year bandwidth scarcity becomes an even greater bottleneck, creating opportunities for companies positioned in optical and interconnect technologies. Hyperscalers are leaning into pluggable optics for flexibility, while co-packaged optics adoption is on the horizon and will drive significant change. Optical components will benefit from 1.6 terabit transceiver ramps linked to NVL72 and Rubin racks. Compute requirements for new AI models are becoming so large they are spilling over what any one datacenter can provide, creating opportunity for long haul optics connecting far flung datacenters together with extreme levels of bandwidth capacity needed.

Skepticism of AI’s durability has kept forward valuations low relative to the growth these companies are experiencing. Semiconductor capital equipment trades in the 30s on forward PE even as it is on the cusp of a step function higher in industry spending. Memory suppliers trade below 10 times, levels suggesting the market believes earnings will soon peak. Networking names trade around thirty times with a wide spread, but earnings growth for some of these companies is over 100%. Nvidia epitomizes AI skepticism – the company has seen revenue growth average over 40% for the past dozen years, yet it trades at 20x estimates for calendar 2026 when it will grow well north of 60%. If AI spending progresses as we expect it will through the decade, merely maintaining multiples at present levels will provide handsome returns on the long side.

AI spending will continue to increase as new technologies—specifically larger models, expanded context windows, and agentic layers—drive enhanced monetization. Fundamentally, the pre-training scaling laws remain intact, a fact recently confirmed by Google’s Gemini 3, which demonstrated that the performance delta between model generations is as significant as ever. This trajectory suggests that the arrival of Blackwell-class models in 2026 will further accelerate demand. Model scale is currently compounding toward seven-trillion-parameter regimes, with training compute growing even faster than parameter counts due to longer runs, multi-modal integration, and the need to impose speed limits to prevent the models from crashing during training. Even as quantization declines toward FP4 and efficiency gains are realized through sparsity, the total token volume processed in both training and inference is rising at an exponential pace.

In 2025, Cursor, an AI coding agent, fundamentally altered the technology landscape by proving that AI agents are commercially viable productivity engines rather than mere research novelties. By leveraging massive context windows to ingest and reason over entire codebases, Cursor transcended simple autocomplete to function as a semi-autonomous engineer. This breakout success validated the agentic thesis, demonstrating that large language models could move beyond passive chat interfaces to execute complex, multi-step workflows with minimal human oversight.

The implications of this shift for infrastructure investment and return on investment have been immediate and profound. Cursor’s ability to maintain state across long sessions demonstrated that effective agents require continuous, stateful compute rather than intermittent processing bursts. This reality directly justifies the heavy capital expenditures we are seeing in high-bandwidth memory, liquid cooling, networking, etc. Furthermore, the ROI calculation has shifted from the cost of generating tokens to the value of completed labor. Because enterprises can now measure ROI in terms of autonomous ticket resolution and code generation, the ceiling for hardware investment has risen dramatically. The efficiency gains of the coding tools far exceed 10% (see Figure 2). But even if efficiency gains are only 10%, the net ROI is over 3,000% ($15K value / $480 AI cost versus the 12 months $150K cost of a software developer). Counter to popular belief, the ROIC of the hyperscalers started to increase after the ChatGPT moment in 2022, commensurate with their sharply increasing infrastructure investments (see Figure 3). They are investing in AI because they see higher returns now; not hoping for higher returns years from now.

We now stand at the precipice of a massive expansion. 2026 may very well be the year that agentic workflows escape the engineering department and permeate every vertical of the enterprise, from finance to legal operations. We are entering the era of “Cowork”—effectively, Cursor for the rest of your work.