2026科技、媒体与电信市场展望报告：把握下一波人工智能浪潮

What’s Inside December 2025 1Introduction CHAPTER 1 2AI Infrastructure andSemiconductors CHAPTER 2 6The Rise of Agentic AI CHAPTER 3 10Media and Content Evolution CHAPTER 4 12Telecom & ConnectivityReinvention CONCLUSION 13What to Expect in theYear Ahead Introduction For AI, 2025 marked a decisive pivot from “modelbets” to “agent bets” as capital flowed toward appliedAI, agentic automation, enterprise integration, andinfrastructure capable of delivering tangible economicROI. While the hype cycle has cooled, and capital hasbecome more selective, funding is robust in areas withclear payback periods. Heading into 2026, investorsand enterprises are anchoring AI deployment aroundmeasurable outcomes, such as unit-cost reduction,cycle time improvement, headcount leverage,productivity, and revenue initiative. Adoption of newAI application layers and AI’s influence on other keyverticals is expected to be front and center in 2026.We view reliability, risk control, complex integration,and compute and infrastructure as key areas of risk todeployment schedules. With the help of the AlphaSense platform, we haveidentified the key emerging themes for technology,media, and telecommunications in 2026. Below, weshare each theme and emphasize areas that have thepotential to inflect in the coming months. CHAPTER 1 AI Infrastructure andSemiconductors The explosive growth of AI workloads — especially large-language models, generative AI and AI inference at scale — ispushing demand for high-performance compute in data centers.Demand for specialty semiconductors, memory, networkinterconnects and packaging technologies is rising in tandem.Hyperscalers projected to increase spend to roughly $500 billionin 2026, which would mark 30% to 35% growth from $367 billionin 2025. Hyperscaler budgets are durable in nature with multiyearvisibility. Spend is set to broaden beyond general-purposeGPUs to custom ASICs, pushing value to leading-edge nodes,advanced packaging, high-bandwidth memory, and high-speedoptics. Meanwhile, power and thermal constraints are pushingliquid cooling technology to become standard in AI datacentersand cloud environments. Demand is also skewing toward cutting-edge process nodes, as modern chips tend to pack cores andlogic to manage memory and bandwidth. As such, fab capacity and foundry supply will remain critical in2026. While cutting-edge technology such as silicon photonicsand co-packaged optics (CPO) are building into designs,pluggable optics will remain the workhorse through 2026. GPUs, TPUs, and ASICs The 2026 cycle is expected to hit a high-water mark for AIsilicon as a share of datacenter capex, with total acceleratorvolume trending well above 2025 levels even as constraintsremain in several critical areas, such as Taiwan Semiconductor’sChip-on-Wafer-on-Substrate (CoWoS) capacity. GPUs fromNvidia are expected to retain market share, but both thecustom ASIC segment and high-attach XPU market are gainingmomentum inside AI server racks.As one expert explains, evenif performance of Google TPUs were to command the level ofNvidia GPU demand, there would still be questions around howmuch TSMC supply would allocate while meeting customerdemand, adding, “It’s not like Google could just print as manyTPUs as there’s demand.” Nvidia reportedits GB300 generation is fully utilized andconfirmed the seven-chip Rubin platform is on track for a2H 2026 ramp. The back half of 2026 is expected to deliveranother step function cost-per-token improvement by raisingpower density, driving architectural changes in the data centermarket and leaning in on next-generation memory and coolingtechnologies.Experts still seehyperscalers rushing to lockin capacity six to 12 months ahead of time, sustaining theinfrastructure investment cycle through 2027. The custom ASIC TAM is projected to be between $30 billionand $35 billion for the top four U.S. cloud service providers(CSP), with Google TPUs capturing the largest share, followedby Amazon Web Services, Meta and Microsoft. Jefferiesbelieves that Google is in a strong position for continued growthand share gain into 2026, owing to the Gemini 3 strength anda potential deal to sell TPUs to Meta. Meanwhile, Amazon isexpected to accelerate scaling of 3nm Trainium, while improvingpower efficiency for its Bedrock and SageMaker workloads. The emerging XPU ecosystem represents specialized computeaccelerators in modern AI infrastructure designed to overcomelimitations of general purpose chips for specific tasks. This is anarea to watch as companies like Broadcom and Marvell connectsystems across racks using optical interconnects. Hyperscalersface a dilemma in balancing tighter software coupling andpotential cost advantages with execution risk, capital intensity,and ecosystem hurdles. HBM Capacity, Pricing, and HBM4 Shift High Bandwidth Memory (HBM) is a foundational enabler for the2026 AI compute cycle. AI hardware suppliers are increasinglyembedding HBM into systems to meet pe