AttoTude is sending radio at nearly a trillion cycles per second down a plain wire - the first interconnect built to stop AI data centers from choking on their own bandwidth.
Inside a test bench in Menlo Park, a length of wire carries a signal oscillating close to one terahertz - a frequency physicists spent decades calling a "gap" because nothing useful lived there. AttoTude moved in. The signal arrives clean four meters away at 224 gigabits per second, and the people watching it are not surprised, which is the most surprising part.
This is a deep-tech company with a narrow obsession: the short, unglamorous links between chips, boards, and servers. Not the GPUs everyone photographs. The cabling between them. It turns out the most expensive problem in artificial intelligence is not making the math faster. It is moving the answers around before they go stale.
The explosive growth of artificial intelligence is constrained by the limits of interconnects.
— AttoTude's founding premiseAbove: a wire doing something wires were politely told they could not. The terahertz gap, now with a tenant.
Every AI cluster is a negotiation between two appetites. The chips want more neighbors to talk to. The wires connecting them - copper - run out of room, throwing away signal and burning power as speeds climb. Optics solve the distance problem but bring cost, heat, and fragile packaging. For a while the industry accepted this as a law of nature.
AttoTude looked at the same data and saw a deadline. By 2029, data centers are projected to add more than 50 million high-end GPUs. Each one needs to talk to thousands of others, instantly, reliably, cheaply. Copper can't stretch that far. Photonics can't get that cheap. Something in between had to exist.
Reliable and cheap - until the data rate climbs, the reach collapses, and the power bill does not.
Fast and far, but it leans on lasers, exotic packaging, and a cost structure hyperscalers wince at.
Pick two of bandwidth, power, and reliability. AttoTude refused to pick.
— The bet underneath the companyThe tension: AI's compute grew like a teenager. Its wiring aged like a filing cabinet.
The fashionable move was to replace wire with light. AttoTude's founders did the opposite - they decided to make plain wire carry radio at terahertz frequencies, borrowing the signaling tricks of fiber optics and running them on standard, high-volume ASICs. No lasers. No custom silicon packaging. Just a frequency band almost nobody had commercialized.
Two people had the resume to attempt it. Dave Welch, co-founder and CEO, helped build the optical-networking company Infinera and holds more than 200 patents. Joy Laskar, co-founder and CTO, has started five companies and spent years as a professor at Georgia Tech, deep in radio-frequency and semiconductor engineering. One knew how to move data with light. The other knew how to bend radio to his will. Terahertz over wire sits exactly where their two careers overlap.
Optical-networking veteran (Infinera co-founder, former CTO of SDL); 200+ patents, 300+ technical publications.
Serial founder of five startups and former Georgia Tech professor specializing in RF and semiconductor systems.
Our THz platform represents a paradigm shift that solves critical bandwidth, power, and reliability challenges.
— Dave Welch, Co-Founder & CEOPictured in spirit: a laser guy and a radio guy walk into a data center. The wire never saw it coming.
A milestone timeline for a company that is younger than most AI demos.
AttoTude forms around THz-over-wire; an ~$11M seed round comes together with Sutter Hill, Canaan, and Wing.
Capital to revolutionize high-speed networking for AI and hyperscale data centers.
Debuts THz radio over wire and demonstrates 224 Gb/s over four meters at 970 GHz.
Led by Mayfield with The Westly Group and existing backers - lifting total funding to roughly $91M.
Adopts Keysight's UXR scope and 89600 VSA/FlexFrame to validate the platform at up to 1 THz and 448G PAM4.
AttoTude's platform is deceptively simple to describe. An ASIC generates and processes terahertz signals. A low-loss wire carries them. Together they push individual lane speeds from 112 Gb/s up toward 448 Gb/s and beyond, over reaches of up to roughly 40 meters - far enough to wire a rack-scale AI cluster end to end without optics.
An ASIC adapted for terahertz frequencies, built on standard, high-volume manufacturing - the part that makes THz affordable.
A low-loss wire that carries terahertz radio chip-to-chip and board-to-board, up to ~40m, no photonics required.
The trick worth pausing on: this rides on standard ASIC processes. That is the whole pitch. Photonics asks the industry to retool. AttoTude asks it to plug something in. One of these is an easier sentence to say to a hyperscaler's procurement team.
Terahertz speed, on the manufacturing line that already exists. The boring part is the breakthrough.
— On why standard ASICs matterReading the chart: the yellow bar is the one keeping data-center architects up at night - in the good way.
Talk is cheap; terahertz instrumentation is not. The first proof point is capital. In April 2025, Mayfield led a $50M Series B, joined by The Westly Group and existing investors Sutter Hill Ventures, Canaan Partners, and Wing Venture Capital. That brought total funding to roughly $91M for a company barely a year old.
The kind of step-function advancement that occurs once in a decade in the semiconductor industry.
— Navin Chaddha, Managing Partner, MayfieldThe second proof point is measurement. In March 2026, Keysight Technologies signed on to validate AttoTude's platform with its UXR oscilloscope and 89600 vector signal analysis software - the same gear used to characterize the bleeding edge of wireless. You do not borrow Keysight's best instruments to measure a whiteboard sketch.
The third is interest. Early evaluations are reportedly underway with major AI infrastructure players, with commercial deployments expected within roughly 12 to 18 months of the raise. Names stay quiet, as they tend to when hyperscalers are kicking the tires on something that could rewire their racks.
Their tools have proven invaluable in validating and evolving our THz technology.
— Dave Welch on the Keysight collaborationFootnote for skeptics: investors fund pitches, but engineers borrow oscilloscopes. The second one is harder to fake.
AttoTude's stated goal is everything-to-everything connectivity: a data center where any chip can reach any other at terahertz speed over cheap wire, so AI compute is no longer gated by how fast data can move. The win condition is invisibility. When the interconnect stops being the thing people complain about, the company has done its job.
There is an environmental edge to this too. Reported figures suggest the approach can move the same throughput in roughly half the physical volume of copper, with better energy efficiency - which matters when data centers are already straining local power grids. Faster and lighter and cooler is a rare trio. AttoTude is betting the trio is real.
The mission, plainly: success looks like silence - the sound of nobody blaming the cables.
Return to that bench in Menlo Park. The wire is still humming near a terahertz, still delivering its signal clean. But now imagine it multiplied across a hall of racks, each link carrying hundreds of gigabits, none of them lasers, none of them exotic, all of them cheap enough that nobody thinks twice. That is the future AttoTude is wiring toward.
The company is young, the customers are unnamed, and commercial deployment is still ahead of it. Skepticism is fair. But the physics has been demonstrated, the capital is real, and the instruments are out. The terahertz gap - that long-empty stretch of spectrum - now has a startup living in it, paying rent in gigabits.
The bottleneck didn't vanish on its own. Somebody had to move into the gap nobody wanted and teach the wire a new language. AttoTude raised its hand.
If the wire learning terahertz is the kind of thing your feed should know about, here are the buttons.
Profile compiled from public sources: AttoTude, Business Wire, FinSMEs, EE Times, Keysight, Crunchbase, and trade press. Figures such as data rates, reach, and funding totals are approximate and drawn from company disclosures and reporting current to 2026. Some financial details (revenue, valuation) are not public. Video links point to searches because no single official demo URL was confirmed.