# 6 Failed Semiconductor Innovations That Found Success in Unexpected Applications - Published: 2026-11-06 - Author: [Semiconductor Magazine](https://semiconductormagazine.com/author/semiconductormagazine-com.md) - Topics: [Q&A](https://semiconductormagazine.com/qa.md) - Source: https://semiconductormagazine.com/qa/6-failed-semiconductor-innovations-that-found-success-in-unexpected-applications/ - Publication: Semiconductor Magazine - Publisher: Connectively (Terkel, Inc.) - Editorial policy: https://semiconductormagazine.com/editorial-policy/ - AI use: The perspectives in this article were written by the named contributors and submitted through Connectively. Connectively's automated system selected them and used AI to write the headline, introduction, section headings, summary sections, and illustration, and to make minor edits for clarity. Articles are assembled and published automatically. Report errors to corrections@connectively.us. ### CUDA Opens Neural Network Training CUDA, Nvidia's 2007 toolkit for programming graphics chips, stayed a niche bet for years before neural network training made it the base of modern AI. Before it came along, researchers had to disguise their maths as graphics work by storing numbers as textures and writing the calculation as a pixel shader, so anyone without graphics programming skills stayed away. Training a neural network is mostly huge batches of matrix multiplication. A GPU's thousands of small cores run that work in parallel far faster than a CPU can. It eventually succeeded because the hardware already suited the maths and CUDA removed the one barrier that kept researchers out. --- ### GaN Chargers Build Market Momentum Gallium nitride power transistors. I remember how GaN was going to replace the silicon MOSFET for power conversion in the 2000s and how that ended. The early GaN transistors were all depletion mode, i.e. normally on. Which means that a failed gate driver results in a short circuit. And with very poor yields to boot. Silly expensive too for what you got. First, the market for enhancement-mode GaN switches "hit a home run" in the little USB-C charger for your phone. The GaN switches are safe enough to use in these chargers. A 65W charger now fits in your pocket - all while using much smaller magnetics than the typical switch-mode power supply. People are willing to pay a lot of money for a small, high-power charger that fits in your pocket. Not for a power supply used in an industrial plant to increase efficiency. Gallium nitride GaN, when it finally hit the market (and worked), found a profitable home in the small, relatively low power, fast production cycle world of USB power supplies. So much so that the volume from this small amount of business was able to drive down the cost per amp, pull the fabs up the learning curve to achieve reasonable yields, and finally have some real data to report on the reliability. Now the volume of production is starting to have an impact in areas that GaN was originally supposed to address. I'm watching this space because of the constraints of power density in server racks. The better the PSUs, the less heat is generated, more compute can be packed into the same space of a rack and it's less likely to have some thermal surprise in the middle of the night. A technology needs to grow up in some cheap, easy to service market before it can make it in some tough market. --- ### Software Unlocks MEMS Mobile Adoption Micro-Electro-Mechanical Systems or MEMS sensors were regarded as an important innovation in semiconductors. They took a long time to mature before they could find a place in cellular technology. Invented in the 1980s, their early tiny mechanical devices in silicon were majorly used in select industrial markets and automotive safety, mainly for airbag triggering. They had been considered a commercial flop for many years as they were costly to produce. Another reason was that soon enough they were seen as useless by the consumers, as there was no software able to process the data obtained from them. The change from the narrow industrial use to the omnipresence of MEMS came when the software had finally caught up to the hardware. When we solve tasks connected with scaling of digital experience and customer operations, we see that the most interesting technological developments are still not being used if they are not incorporated into a recent platform. GPS, gyroscope and other technology used in the mobile devices had transformed from just safety functions to means of communication during use. The success of this invention was achieved due to turning what was supposed to be a stand-alone device into a platform-compatible data source. This change shows that the success in semiconductor technology does not totally depend on the chip itself. In fact, it depends on the possibility of the hardware-software pairing. The lesson for business is that sometimes the innovation is regarded a bust simply because the current technological background is not ready for this type of devices. --- ### Space Validates Silicon Solar Cells One of the most interesting semiconductor innovations that struggled commercially before finding an unexpected market was the silicon solar cell. Bell Labs introduced the first practical silicon solar cell in 1954, but manufacturing costs made widespread terrestrial use economically impractical. Its breakthrough came in space, where reliability and continuous power generation mattered more than the initial purchase price. The 1958 Vanguard 1 satellite demonstrated that solar cells could power equipment for years, establishing a valuable new application for semiconductor technology. What makes this example compelling is that the technology did not need an entirely new purpose; it needed a market with different economic priorities. On Earth, solar cells initially competed against cheaper conventional power sources. In space, replacing batteries was difficult or impossible, making long-term power generation significantly more valuable. The same technology that struggled to justify its cost in one environment offered a clear operational advantage in another. I see an important commercial lesson here for evaluating emerging technologies. I recommend assessing innovation through three factors: the problem it solves, the cost of existing alternatives, and the value of its unique capabilities. A semiconductor technology that appears commercially unsuccessful in a price-sensitive market may become essential in an application where reliability, efficiency, or operating conditions carry greater weight. The lesson is that technological success often comes from finding the right application before achieving the lowest price. --- ### Graphics Processors Find a Killer Workload The answer here is the GPU, and it's not even close. NVIDIA built the GPU to render triangles faster for video games. That was the entire point. For years, the broader tech industry looked at GPUs as niche silicon for gamers and hobbyists. Not serious compute. Not enterprise-grade. Wall Street didn't care. Enterprise IT departments didn't care. The GPU was a "toy" in the same way the internet was a "toy" in 1994. Then around 2012, a researcher named Alex Krizhevsky used NVIDIA GPUs to train a deep neural network called AlexNet, and it obliterated the competition in an image recognition contest by a margin nobody had ever seen. That single moment cracked open the door. Researchers realized that the massively parallel architecture designed to shade pixels in Call of Duty was almost perfectly suited for the matrix math at the heart of machine learning. A chip built for entertainment became the backbone of artificial intelligence. Jensen Huang saw it before almost anyone else in the industry. He pivoted NVIDIA's roadmap, invested in CUDA, built an entire software ecosystem around general-purpose GPU computing, and bet the company on a future that hadn't arrived yet. Today NVIDIA is one of the most valuable companies on Earth, and every major AI model, including the ones powering Magic Hour, runs on their hardware. The reason the GPU eventually succeeded in AI comes down to one thing: the architecture was right, it was just waiting for the right problem. Parallel processing at massive scale turned out to be exactly what neural networks needed. The market didn't exist when the chip was designed. The chip created the market. I think about this constantly when building Magic Hour. The tools we use today were not built for what we're using them for. Open-source video models were research projects 18 months ago. Now millions of people use them through our platform to create content that drives real revenue. The lesson is simple: the best innovations don't succeed because the original use case wins. They succeed because someone finds the use case the inventor never imagined. --- ### CCDs Transform From Memory to Imaging Hi! A good example is the charge-coupled device (CCD)—a semiconductor technology conceived for computer memory that found its greatest success in digital cameras and astronomy. Willard Boyle and George Smith invented it at Bell Labs in 1969. CCDs store and move tiny packets of electrical charge, but as memory they had drawbacks: retrieving information required shifting charges through a sequence, which slowed access, and some charge could become trapped along the way. IEEE's oral history describes why that approach struggled. The unexpected opportunity was image sensing. Light striking silicon generates charge, so each collection site could represent a pixel: brighter light produces more charge. Michael Tompsett and colleagues developed this into practical imaging devices, helping transform photography and astronomical observation. I think it succeeded because the new application suited its underlying physics. Reading pixels sequentially works well for images, while high light sensitivity and compact, durable construction offered real advantages over earlier camera tubes. Further engineering—including cooling sensors for long astronomical exposures—made those advantages usable. One nuance: researchers recognized imaging possibilities early. This was a failure to establish a successful memory business followed by success in another application, rather than a forgotten invention rediscovered decades later. ## Related - [7 Industries Revolutionized by Semiconductor Technology](https://semiconductormagazine.com/qa/7-industries-revolutionized-by-semiconductor-technology.md) - [6 Small Improvements in Semiconductor Technology With Big Impacts](https://semiconductormagazine.com/qa/6-small-improvements-in-semiconductor-technology-with-big-impacts.md) - [6 Barriers to Wider Adoption of Emerging Semiconductor Technologies](https://semiconductormagazine.com/qa/6-barriers-to-wider-adoption-of-emerging-semiconductor-technologies.md) - [How Semiconductor Teams Resolve Priority Conflicts](https://semiconductormagazine.com/qa/how-semiconductor-teams-resolve-priority-conflicts.md)