[{"data":1,"prerenderedAt":24},["ShallowReactive",2],{"article-silicio-memoria-ai-locale-robotica-en":3},{"slug":4,"title":5,"description":6,"date":7,"readTime":8,"author":9,"image":10,"macro":11,"categories":12,"subcategories":16,"tags":18,"content":23},"silicio-memoria-ai-locale-robotica","Local AI and Robotics in 2026: Edge NPUs, Memory, and Shared Infrastructure","The article on robotics asked what infrastructure would become indispensable. The answer is the same one that explains the shift toward smaller models: edge NPUs and memory, shared between robots and local AI.","2026-08-20",11,"Punta.dev","/images/articles/silicio-memoria-ai-locale-robotica/cover.webp","tech",[13,14,15],"ai","hardware","robotica",[17],"innovazione",[19,20,21,22],"Artificial Intelligence","Investments","Robotics","Hardware","\u003Cdiv class='article-intro'>\u003Cp>\u003Cem>The article\u003Ca href='/blog/guerra-silenziosa-robot-ai-fisica' class='underline decoration-purple-400 hover:text-purple-200'> on robotics\u003C/a> ended with a question left deliberately open: What infrastructure will become essential for robots to truly function on a large scale? In the article\u003Ca href='/blog/intelligenza-artificiale-locale-modelli-piccoli' class='underline decoration-purple-400 hover:text-purple-200'> on small models\u003C/a>, we explained why AI is learning to operate closer to us—on the device rather than in the cloud. These two pieces share the same answer, and it’s worth bringing them together.\u003C/em>\u003C/p>\u003C/div>\u003Cp>A robot that needs to avoid an obstacle, or a phone that needs to understand a voice command, faces the exact same constraint: they cannot afford to wait for a round-trip to a distant server. \u003Cstrong>Local\u003C/strong>, real-time, low-power inference is needed. It’s the same problem, with two different sides—and the companies producing the silicon to solve it are becoming an interesting vantage point for both topics together.\u003C/p>\u003Ch2>Why Robotics and Local AI Require On-Device Inference and Low Latency\u003C/h2>\u003Cp>In the article on robotics, the central point wasn’t the robot itself, but everything needed for it to truly function: physical data, simulation, and motor control. There’s one aspect that article didn’t explore in depth: a robot’s motor control and sensory perception must run \u003Cstrong>on the device itself\u003C/strong>, not in the cloud—the latency of a network round-trip would be incompatible with avoiding an obstacle in real time, and connectivity is never 100% guaranteed in a factory or warehouse.\u003C/p>\u003Cp>This is exactly the principle of “right-sizing” described in the article on small models, applied to a context where the stakes are physical rather than merely computational: you need a model small enough to run on-device, yet capable enough to perceive and react in real time.\u003C/p>\u003Ch2>Edge NPUs: The Chips Powering Local AI, Robotics, and Smart Devices\u003C/h2>\u003Cp>Manufacturers of edge NPUs (Neural Processing Units) are explicitly positioning these chips to cover both “phone/PC” and “industrial robot” use cases with the same underlying architecture.\u003C/p>\u003Cfigure class='my-10'>\u003Cdiv class='bg-gray-800/40 backdrop-blur rounded-2xl p-6 md:p-8'>\u003Cp class='text-white font-semibold text-lg mb-1'>Qualcomm: A Declared Leader in Edge AI, Priced Like a Cyclical Stock\u003C/p>\u003Cp class='text-gray-400 text-sm mb-6'>Forward P/E Comparison (Qualcomm: stated range 10.5–16x, midpoint shown)\u003C/p>\u003Cdiv class='relative' style='height:220px'>\u003Cdiv class='absolute inset-x-0 border-t border-gray-500/20' style='top:75.0%'>\u003C/div>\u003Cdiv class='absolute inset-x-0 border-t border-gray-500/20' style='top:50.0%'>\u003C/div>\u003Cdiv class='absolute inset-x-0 border-t border-gray-500/20' style='top:25.0%'>\u003C/div>\u003Cdiv class='absolute inset-0 flex items-end justify-center gap-8 sm:gap-16'>\u003Cdiv class='flex flex-col items-center w-20 sm:w-24 flex-shrink-0'>\u003Cspan class='text-white font-bold mb-2 text-xs sm:text-sm text-center whitespace-nowrap'>10.5–16×\u003C/span>\u003Cdiv style='height:105px' class='w-full rounded-t-lg bg-gradient-to-t from-purple-600 to-purple-400 border-t-2 border-white/30 shadow-[0_0_14px_rgba(192,132,252,0.35)]'>\u003C/div>\u003Cspan class='text-gray-300 text-[10px] sm:text-xs mt-2 sm:mt-3 text-center leading-tight'>Qualcomm\u003Cbr>(edge AI, robotics, automotive)\u003C/span>\u003C/div>\u003Cdiv class='flex flex-col items-center w-20 sm:w-24 flex-shrink-0'>\u003Cspan class='text-white font-bold mb-2 text-xs sm:text-sm text-center whitespace-nowrap'>~22×\u003C/span>\u003Cdiv style='height:220px' class='w-full rounded-t-lg bg-gradient-to-t from-pink-600 to-pink-400 border-t-2 border-white/30 shadow-[0_0_14px_rgba(192,132,252,0.35)]'>\u003C/div>\u003Cspan class='text-gray-300 text-[10px] sm:text-xs mt-2 sm:mt-3 text-center leading-tight'>Nvidia\u003Cbr>(data centers)\u003C/span>\u003C/div>\u003C/div>\u003C/div>\u003Cp class='text-gray-500 text-xs mt-6 text-center italic'>Source: market data, Qualcomm Investor Day, June 2026\u003C/p>\u003C/div>\u003C/figure>\u003Ch3>Qualcomm: From Mobile to Edge AI, Covering Robotics, Automotive, and AI PCs\u003C/h3>\u003Cp>Qualcomm trades at a forward P/E between 10.5 and 16 times, less than half of Nvidia’s 22x—despite the fact that its 2026 Investor Day explicitly identified “robotics” and “physical AI” as one of the new areas of \u003Cspan class='glossary-term' onclick=\"toggleGlossaryTerm(event, this)\">\u003Cspan class='glossary-popup'>TAM (Total Addressable Market): the total potential market that a company could serve with its product—a measure of how much it can grow, not of how much it earns today. \u003Cspan class='glossary-close' onclick=\"toggleGlossaryTerm(event, this)\">Close ✕\u003C/span>\u003C/span>\u003C/span>, along with automotive and AI PCs, for a combined total addressable market estimated at approximately $1,700 billion by 2030. The market still largely values it as a supplier of smartphone chips, not as structural AI infrastructure that also spans the physical world.\u003C/p>\u003Ch3>CEVA: NPU Licensing and Intellectual Property for Edge AI\u003C/h3>\u003Cp>\u003Cstrong>CEVA\u003C/strong>, an intellectual property licensor for edge NPUs that is much smaller than Qualcomm, now explicitly uses the term “Physical AI” in its communications to investors—the same thematic area as the article on robotics. Here, rather than a single aggregate figure, it’s worth looking at the actual reports from the investment banks: on May 12, 2026, following the quarterly earnings report, TD Cowen raised its price target from $24 to $45, UBS from $42 to $48, Oppenheimer from $30 to $42, Rosenblatt from $40 to $45, and Stifel set a target of $42—all with a “Buy” rating. However, this is not a unanimous consensus: \u003Cstrong>JPMorgan\u003C/strong>, which initiated coverage on May 8, set a target price of just $30 with a “Neutral” rating—the most cautious of the group. The 12-month consensus, as of August 2026, stands at around $47, compared to a price that has fluctuated between $28 and $40 over the same period depending on the moment—a historically very volatile stock (12-month range: $17.02–$51.25). The risk to keep in mind remains the same: CEVA is still reporting a GAAP loss (negative P/E), so any target price reflects a bet on future royalties, not profits already realized.\u003C/p>\u003Ch2>LPDDR and HBM Memory: Why AI Growth Is Driving Up Costs Even at the Edge\u003C/h2>\u003Cp>There is a more subtle—and more direct—connection to the article on small models. Both a robot and a local inference project like \u003Ca href='/blog/intelligenza-artificiale-locale-modelli-piccoli' class='underline decoration-purple-400 hover:text-purple-200'>DwarfStar4\u003C/a> require abundant, fast RAM close to the chip. But the same rush toward HBM for data centers, as described in the article on bottlenecks, is also draining this “consumer-grade” memory: LPDDR, the type found in phones, PCs, and embedded systems for robotics.\u003C/p>\u003Cfigure class='my-10'>\u003Cdiv class='bg-gray-800/40 backdrop-blur rounded-2xl p-6 md:p-8'>\u003Cp class='text-white font-semibold text-lg mb-1'>The price of memory for local AI skyrocketed in the same quarter\u003C/p>\u003Cp class='text-gray-400 text-sm mb-6'>LPDDR5X Price Index (base 100 = start of quarter)\u003C/p>\u003Cdiv class='relative' style='height:220px'>\u003Cdiv class='absolute inset-x-0 border-t border-gray-500/20' style='top:75.0%'>\u003C/div>\u003Cdiv class='absolute inset-x-0 border-t border-gray-500/20' style='top:50.0%'>\u003C/div>\u003Cdiv class='absolute inset-x-0 border-t border-gray-500/20' style='top:25.0%'>\u003C/div>\u003Cdiv class='absolute inset-0 flex items-end justify-center gap-8 sm:gap-16'>\u003Cdiv class='flex flex-col items-center w-20 sm:w-24 flex-shrink-0'>\u003Cspan class='text-white font-bold mb-2 text-xs sm:text-sm text-center whitespace-nowrap'>100\u003C/span>\u003Cdiv style='height:97px' class='w-full rounded-t-lg bg-gradient-to-t from-purple-600 to-purple-400 border-t-2 border-white/30 shadow-[0_0_14px_rgba(192,132,252,0.35)]'>\u003C/div>\u003Cspan class='text-gray-300 text-[10px] sm:text-xs mt-2 sm:mt-3 text-center leading-tight'>Start of Q2 2026\u003C/span>\u003C/div>\u003Cdiv class='flex flex-col items-center w-20 sm:w-24 flex-shrink-0'>\u003Cspan class='text-white font-bold mb-2 text-xs sm:text-sm text-center whitespace-nowrap'>+78–83%\u003Cbr>(≈181)\u003C/span>\u003Cdiv style='height:220px' class='w-full rounded-t-lg bg-gradient-to-t from-pink-600 to-pink-400 border-t-2 border-white/30 shadow-[0_0_14px_rgba(192,132,252,0.35)]'>\u003C/div>\u003Cspan class='text-gray-300 text-[10px] sm:text-xs mt-2 sm:mt-3 text-center leading-tight'>End of Q2 2026\u003C/span>\u003C/div>\u003C/div>\u003C/div>\u003Cp class='text-gray-500 text-xs mt-6 text-center italic'>Source: TrendForce, May 2026 survey\u003C/p>\u003C/div>\u003C/figure>\u003Cblockquote>Memory currently accounts for 30–40% of a smartphone’s component costs, compared to the historical 10–15%. The exact same mechanism applies to every embedded system—including robots—that requires ample RAM close to the chip to process perception and control locally.\u003C/blockquote>\u003Cp>It’s the same Jevons paradox described in the article\u003Ca href='/blog/intelligenza-artificiale-locale-modelli-piccoli' class='underline decoration-purple-400 hover:text-purple-200'> on small models\u003C/a>, applied to hardware rather than energy: local AI should make everything cheaper and more efficient, but the aggregate demand required to enable it—more chips, more memory, across more devices—is driving up the price of the components needed to make it happen, for everyone, not just for those building data centers.\u003C/p>\u003Ch2>Investing in Edge AI: Where Value Is Concentrated Among NPUs, Memory, and Robotics\u003C/h2>\u003Cp>The question left open in the article\u003Ca href='/blog/guerra-silenziosa-robot-ai-fisica' class='underline decoration-purple-400 hover:text-purple-200'> on robotics\u003C/a>—where the value will truly be concentrated—finds a partial but concrete answer here: not only in the finished robot or the most elegant model, but in \u003Cspan class='glossary-term' onclick=\"toggleGlossaryTerm(event, this)\">\u003Cspan class='glossary-popup'>edge siliconEdge\u003C/span> silicon\u003Cspan class='glossary-popup'>: chips designed to perform artificial intelligence calculations directly on the device—phone, robot, PC—rather than in a remote data center.\u003Cspan class='glossary-close' onclick=\"toggleGlossaryTerm(event, this)\">Close ✕\u003C/span>\u003C/span>\u003C/span> which allows both to perceive and react without relying on the cloud, and in the memory that this silicon requires to function. It’s an infrastructure shared between two narratives that seem worlds apart—a phone summarizing an email locally and a robotic arm avoiding an obstacle in a factory—but which are literally competing for the same production capacity of chips and memory.\u003C/p>\u003Ch2>Investment Risks: Cyclicality, Litigation, and Execution in Edge AI\u003C/h2>\u003Cp>It’s worth taking a closer look here, because each of these names carries a concrete, long-standing risk—not a generic one.\u003C/p>\u003Ch3>DRAM and LPDDR Memory: The Cyclical Risk of the Semiconductor Industry\u003C/h3>\u003Cp>In 2022–2023, during the last cycle reversal, \u003Cstrong>SK Hynix posted a net margin of approximately -28% for the entire year\u003C/strong>—one of the world’s leading memory manufacturers losing money on nearly every dollar of sales—while Micron’s stock had lost about half its value from its 2022 peak. Some analysts now estimate that DRAM prices could fall by 80–90% over the next three years, when new production capacity (including Chinese manufacturers CXMT and YMTC, which have already captured 8% and 13% of their respective global markets) comes online. Even investors known for anticipating past bubbles, such as Michael Burry, have publicly taken a bearish stance on Micron based precisely on this premise. Morningstar, in a recent report, explicitly stated that the slowdown in the AI cycle could occur “as early as 2026.” None of these signals suggest this will happen immediately—the analysts themselves note that this cycle has structural characteristics different from previous ones—but the sector’s history argues for caution, not certainty.\u003C/p>\u003Ch3>Qualcomm vs. ARM: The October 5, 2026, Trial and the Risk to Edge AI\u003C/h3>\u003Cp>Qualcomm has already won a lawsuit against ARM in 2024–2025 regarding the use of Nuvia technology in its CPU cores. But there is a \u003Cstrong>second, separate lawsuit\u003C/strong> in which Qualcomm itself is accusing ARM of breach of contract regarding licensing fees—and the trial, according to the most recent SEC filings, is scheduled for \u003Cstrong>October\u003C/strong> 5, \u003Cstrong>2026\u003C/strong>. ARM states in its financial statements that Qualcomm accounts for 9% of its total revenue and warns that the outcome is uncertain and could require “significant legal expenses” regardless of the result. This is a concrete, near-term catalyst, not an abstract risk.\u003C/p>\u003Ch3>CEVA: The Execution Risk Between NPU Licensing and Future Royalties\u003C/h3>\u003Cp>It bears repeating clearly: the fact that six out of seven banks have a target price higher than the current price guarantees nothing. JPMorgan, the most cautious bank in the group, remains at “Neutral.” CEVA has not yet demonstrated its ability to convert signed licensing agreements into recurring royalties based on actual volumes—this is exactly the kind of bet that may not materialize, or may materialize much more slowly than analysts’ targets imply.\u003C/p>\u003Cp>None of these three stories is therefore a sure bet for investors. But this is the right question to keep in mind, viewing these three articles as a single thread rather than three separate stories—and keeping the risks just as much in view as the opportunities.\u003C/p>\u003Cdiv class='glossary'>\u003Cp>\u003Cstrong>Qualcomm\u003C/strong> — Investor Day 2026 (June 24, 2026); Form 10-Q filed with the SEC (details on the ARM lawsuit, trial set for October 5, 2026)\u003C/p>\u003Cp>\u003Cstrong>ARM Holdings\u003C/strong> — Form 20-F/6-K filed with the SEC (9% of revenue exposed to Qualcomm, disclosed legal risk)\u003C/p>\u003Cp>\u003Cstrong>CEVA\u003C/strong> — quarterly press releases; target prices from individual banks (TD Cowen, UBS, Oppenheimer, Rosenblatt, Stifel, JPMorgan) via TipRanks and StockAnalysis, May–August 2026\u003C/p>\u003Cp>\u003Cstrong>TrendForce\u003C/strong> — LPDDR5X/LPDDR4X price survey, May 2026\u003C/p>\u003Cp>\u003Cstrong>SK Hynix, Micron\u003C/strong> — historical margin and pricing data for the 2022–2023 cycle (Motley Fool analysis, July 2026)\u003C/p>\u003Cp>\u003Cstrong>Morningstar\u003C/strong> — research note on the cyclical nature of the semiconductor industry (2026)\u003C/p>\u003Cp>\u003Cstrong>Financial Times, Reuters, International Federation of Robotics\u003C/strong> — sources cited in the linked article on robotics\u003C/p>\u003C/div>",1787586114447]