Most Companies Use AI. Few Run on It. Phison Wants to Close the Gap

"Everyone wants AI in their business. But getting AI to actually do the work? That's a different story." With those words Michael Wu, president and general manager of Phison US, opened a LinkedIn post on October 18, and they are a fair description of where most companies stand in late 2026. At TechCrunch Disrupt 2026 in San Francisco, his colleague Michelle Chan put figures on the issue: 76 percent of businesses are using AI, but only 14 percent have fully integrated it into their core operations. For a Taiwanese firm whose claim to fame is the flash controllers inside laptops and servers, that gap is also a business plan.

Most companies are using AI, but few are running on it

Wu could not resist a joke about the figures. "Apparently, AI is getting hired faster than it's getting onboarded!" he wrote. It is something most managers have witnessed: a team signs up for a chatbot, a pilot gets a slide in the quarterly review, and nothing changes in how orders are processed.

Speaking from the Breakout Stage at Moscone West, Chan put it another way. "The biggest challenge in AI is no longer building the model," she said. Many organizations have done some experimenting, but very few have made AI part of the work that runs the business. Surveys may not agree on what "integrated" means, but the trend is evident.

In his post, Wu argued that the next phase "isn't about who adopts it first, but who can turn it into real business value." He was also clear about the obstacles. "Data privacy, infrastructure costs, and operational reliability remain major challenges. That's why I believe local and on-premises AI will play an increasingly important role." A factory is not going to halt production over a cloud service failure. A bank or a hospital will not let a model run with its sensitive records wherever it likes. In those businesses AI has to earn its keep or remain a demo.

The real bottleneck is memory, not the model

Phison's approach starts with a problem most business leaders never hear about. Large language models are memory hungry, and the fast memory on a graphics processor runs out well before its compute does. Add a long contract or a larger model, and a machine that performed fine in testing hits a wall.

Memory is Phison's specialty. The company says its controllers power more than one in every five SSDs shipped worldwide. To get around the memory wall it offers Pascari aiDAPTIV, a combination of aiDAPTIV Middleware and dedicated flash drives it calls aiDAPTIV Cache Memory. The system keeps the most active parts of an AI workload in GPU memory, stages the next layer in system memory, and parks less active data on flash until it is called for.

"As AI models grow into tens and hundreds of billions of parameters, the industry keeps hitting the same wall with GPU memory limitations," Wu said when Phison expanded the technology at CES in January. "By expanding GPU memory with high-capacity, flash-based architecture in aiDAPTIV+, we offer everyone, from consumers and SMBs to large enterprises, the ability to train and run large-scale models on affordable hardware."

Phison does not pretend flash matches DRAM for speed. The calculation is that an AI system holds plenty in memory that is not time sensitive, and that work can be handled by cheaper flash.

What Phison actually sells to make a business AI capable

In its COMPUTEX announcement in June, Phison put the entire stack on display. At the head of it is the new Phison AI Data Platform, which combines infrastructure hardware, resource orchestration, AI software modules and application services so a company can set up its own local AI environment.

At the machine level, the Pascari aiDAPTIV AI20EH drive won a COMPUTEX Best Choice Award as an AI PC solution. In Phison's internal tests on identical AI PC hardware, the drive improved inference performance by up to 102 times, cut memory use by 67 percent and lowered local deployment costs by up to 53 percent. Those vendor figures deserve independent verification. To put a dent in cloud bills, the Phison Hybrid Router keeps suitable AI work on the local PC, and a tie-up with Intel pairs aiDAPTIV with Intel Core Ultra Series 3 processors.

In the data center, the Pascari D206V enterprise SSD offers as much as 245.76 terabytes on one drive and took a COMPUTEX Best Choice Golden Award. CEO K.S. Pua framed the shift plainly: "In the AI era, competitiveness is no longer defined solely by compute performance, but by data access efficiency and system architecture integration capabilities."

A practical path for companies that want AI to do real work

A business owner would do well to treat AI as infrastructure rather than a subscription. Single out one workflow that matters, whether support tickets, quality reports or contract review, and decide where its data should live. If that data is sensitive, running the model on an on premises server, an AI PC or even a workstation removes a whole class of risk.

It also means sizing memory, not just buying more GPUs. Wu made that case at NVIDIA GTC in March, saying Phison's approach "enables larger models and long-context inference on local AI platforms without requiring additional GPU hardware, helping organizations keep AI workloads local while planning infrastructure investments more predictably."

Phison is no stranger to this way of working. Chan told the TechCrunch crowd that more than 72 percent of Phison's staff use AI every day, saving some 8,400 work hours a week, and that the company's platform is in use at more than 30 hospitals, 10 banks and about a dozen large enterprises.

Corporations looking to make their products or operations AI capable with Phison's technology should start with Wu. He leads Phison in the United States, has been the voice of its local AI push all year, and can be reached through his LinkedIn profile. Teams that want to test the technology first can use the "Discuss an integration" option on Phison's inference page.

How startups can build with Phison instead of around it

There are two clear paths for a startup. Software teams building AI applications can make aiDAPTIV part of their design from the outset, so a product that needs a large model can run on a customer's own PC or edge device rather than on a rented GPU in the cloud. For a startup with hospitals, banks or factories as clients, that can be what turns a pilot into a contract.

Hardware startups can go further. Phison's IMAGIN+ design services put custom flash controllers, firmware tuning and validation testing at their disposal, along with shared R&D labs. Phison says there is no set minimum order quantity or engineering fee; a project starts with the customer's product concept, a rough estimate of cost and timeline, and a scope of work. For a young company building a robot or an edge AI box, that is engineering depth it could never put on the payroll.

Phison's invitation from the TechCrunch Disrupt stage

Chan's October 13 talk at TechCrunch Disrupt carried the most pointed message for the founders in the room. Speaking in the session "Taiwan Powering the Next AI Revolution: Connect, Invest & Scale," she began with Phison storage technology that made its way to the Moon inside a hardware data center. Phison is growing its corporate venture capital arm and invited startups to pursue joint development and strategic investment with the company, particularly those building on its AI platform.

Phison's Silicon Valley AI Demo Site is another place where a startup can put its product on display and get in front of prospective customers. Phison did not put numbers on its fund or share portfolio details, but a strategic investor can offer what a purely financial one cannot: the hardware, the manufacturing ties and the firmware engineers. The session summed it up in one line: "Silicon Valley invents. Taiwan builds."

Will the companies that win with AI be the ones that keep it close?

If Phison is right, business AI in the coming years will look less like a single giant cloud and more like a network of capable local machines. An accounting firm could run a private assistant on an office workstation. A clinic could put together patient summaries without the data leaving the building. A factory's AI would not go idle when the internet goes down. For Silicon Valley founders, that opens a market for AI software that does not need top tier hardware to run, and for workers it could turn AI from a side project into part of the everyday routine. If the 14 percent integration rate climbs toward a majority, the companies supplying the plumbing will have done much of the work.

But there are real concerns. Flash memory is slower than DRAM, so some workloads will always need costly fast memory. Phison's performance and cost numbers come from its own testing, and independent benchmarks from real deployments are still thin. Memory and flash prices also move in cycles, so the cost case for a local setup can shift from year to year. Local AI puts the onus on the business to patch models, secure machines and keep hardware current. Smaller firms may lack that talent, which is the gap Phison and its partners set out to fill.

Then there is the debate. Cloud providers argue that their scale, constant model updates and pay as you go pricing make them the safer choice, and many companies will settle on a hybrid of the two. Advocates of local AI counter that per use pricing, data sovereignty rules and outage risk all point the other way. What neither side disputes is that the model is no longer the difficult part. It is the data, the people and the plumbing.

What to watch: independent tests of aiDAPTIV on AI PCs and workstations as they ship, the first startup investments Phison's venture arm makes public, the companies that show up at its Silicon Valley AI Demo Site, and what enterprise surveys say about full integration a year from now. A rise in that 14 percent figure would signal that the shift from AI experiments to AI operations has begun.

Sparknify coordinated the TechCrunch Disrupt session where Chan spoke, and our report on Taiwan's moment on the Disrupt stage has more on ITRI, Phison and the startups that joined them.

The AI race inside companies is no longer about who signs up first, but about who makes it do the work.

Subscribe to Sparknify for Silicon Valley tech news and the people behind it.