Nigel Toon has spent almost three decades in the semiconductor industry, accumulating multiple patents and selling two start-ups, including one to Nvidia for about $400 million.

None of that mattered three years ago when Toon was trying to raise money for XMOS, a chipmaker for digital electronics.

Venture capitalists wouldn’t touch silicon. Start-ups required too much capital to get off the ground and had a brutally high mortality rate. On the occasion when Toon’s pitch excited a particular venture investor, the rest of the firm balked.

“They’d go to the partner meeting on Monday morning and the other partners would fall on the floor laughing,” said Toon, in a recent interview with CNBC. “They’d say, `You want to do what?’ It was almost toxic.”

Fast forward to 2017 and there’s no more laughing.

Toon is now co-founder and CEO of Graphcore, which spun out of XMOS in mid-2016. The U.K.-based company is developing silicon for high-performance servers to run artificial intelligence workloads. It’s playing right into one of the tech industry’s hottest trends and reminding Silicon Valley where it got its name.

Semiconductor start-ups are poised to raise $1.6 billion this year, up from $1.3 billion in 2016 and $820 million in 2015, according to CB Insights. In particular, chipmakers targeting AI are seeing a surge of interest as investors see a massive math problem staring them in the face.

The computational requirements for training machines — including many of the 20 billion connected devices expected to be in circulation by 2020 — can’t be satisfied by existing processors from Intel and Qualcomm or even by Nvidia’s graphics processing units (GPUs). There’s a whole new era of computing on the horizon that requires processing vision, voice and motion data while simultaneously conserving power.

“AI is like a nitro booster for these next-gen silicon companies,” said Paul Holland, a partner at Foundation Capital, which invested in Graphcore’s $30 million round in October.

In the pre-booster days, life was a lot more difficult. Chipmaker GCT Semiconductor was on file to go public in 2012 but never made it out and withdrew two years later due to a lack of investor demand. Audience, a maker of audio processors, was acquired in 2015 for $85 million, or more 70 percent below its IPO price from three years earlier.

XMOS was able to raise money from strategic investors in 2014, but Toon said it took six months to close the round even though the company had a growing business. There were far more start-ups that couldn’t find continuing investment and faded into oblivion.

Today’s emerging players, by contrast, are talking a big game.

Graphcore claims it can improve performance in the cloud and data center by 10 to 100 times. Its story is compelling enough to investors that immediately following a $30 million round in October, the company was inundated with interest from potential backers and reeled in another $30 million in July, led by London-based venture firm Atomico. Graphcore is in its final phases of product development and has lined up early customers.

“We weren’t planning to raise more capital until next year,” Toon said. “Now we’re in a position to accelerate our growth.”

Many of the new silicon start-ups are still in stealth mode in an effort to stay below the radar of the industry incumbents.

Cerebras Systems has raised over $100 million and is valued at $860 million, according to a Forbes story in August, citing data from PitchBook. While the company has kept silent about its products, its website lists nine open positions, including a microcoder to “create high-performance linear-algebra and machine-learning kernels for custom processors.”

Similarly, don’t expect any information from the website — or landing page — for Groq, a chip start-up founded last year by two ex-Googlers and funded with $10 million from Social Capital. Even in stealth, the company caught the attention of industry veteran Krishna Rangasayee, who left Xilinx last month after 18 years to join Groq as operating chief. Rangasayee declined to be interviewed for this story and said the company is “heads down right now.”

As with any buzzy start-up trend, this one certainly could fizzle before generating meaningful returns.

Shahin Farshchi, a Mythic investor and partner at Lux Capital, is one of the few venture capitalists who’s already made money in the latest silicon boom. He backed Nervana Systems, which Intel bought for over $400 million last year to bring machine learning to the data center.

“Everything else may be a smoking crater,” said Farshchi, who’s also an investor in autonomous driving start-up Zoox. But that’s not what he expects to happen. “My hypothesis here is that there’s a need for a fundamentally new class of products,” he said.

There are good reasons for Farshchi to be optimistic.

The number of computing devices that require cutting-edge processors to run AI workloads is exploding at home, in the car, on the phone and in the data center. That means potentially billions of gadgets, big and small, demanding performance many times more powerful than what’s available today.

“Looking around the room right now, I see 15 different devices that could use an AI brain inside,” said Mike Henry, Mythic’s CEO, who founded the company in 2012 after finishing his PhD in electrical and computer engineering.

Henry, speaking from a conference room in Mythic’s headquarters in Redwood City, California, described his early years toiling away in Blacksburg, Virginia, home to his alma mater Virginia Tech. While teaching college courses on microchip design, Henry had had to scrape for government funding, primarily from the Air Force Academy, to pursue independent work on low-powered mobile technology.

“No VC would’ve funded it back then,” said Henry.

But by 2016, every venture firm had a thesis around AI, with partners blogging about machine learning, deep learning and neural networks.

“Before we knew it, we had four competitive term sheets,” Henry said. “It was like a light switch.”

Mythic raised $15 million in two rounds over a seven-month period. Henry said the company is currently closing a debt round to fund manufacturing and will likely bring in more equity sooner than it previously expected. He expects device makers to test the company’s chips this year and the product to hit the market in 2018.

Engineers at Carnegie Mellon University in Pittsburgh, home to the country’s top AI program, are well aware of how much the world needs innovations in silicon.

The machine learning department is performing research on brain image analysis, protein localization in cells and understanding complex biological systems. To run those projects, Nvidia’s GPUs were eating up so much power across campus that the school had to limit their use.

“The problem with GPUs is they’re very power intensive,” said Andrew Moore, dean of Carnegie Mellon’s computer science school. “We had to put a moratorium on people putting big GPU clusters in their offices.”

The software developers that surround Moore need more efficient hardware that “doesn’t require us to build power stations next to the data center,” he said.

–CNBC’s Lora Kolodny contributed to this report