She might not be a fan of surprises, but it turns out that some of the most important events in Dr Angelina Totović’s career have been shaped by exactly those. As a Principal Engineer at Marvell, she is working at the AI frontier of photonics. She trusts the one thing that always seems to behave: physics.
Principal Engineer at Marvell
The frontier of photonics is among the least predictable corners of modern engineering, a place where quiet research can grow into $5.5 billion dollar acquisitions inside a five-year window, and where the engineers closest to the science are often so invested in the technological developments that they are the last to learn how much value they have actually created. Angelina has been inside that turbulence from the start. The deal Marvell struck with Celestial AI in late 2025, the largest event in her professional life so far, arrived through her newsfeed the same way it arrived for everyone else. The role she holds today did not exist when she first sat down to work on the technology that produced it. Read across the whole arc, her career has been one long exercise in learning to work with things she could not predict.
Over Teams from Thessaloniki, an hour borrowed from a calendar that does not have many to spare, she begins where she always begins. With the physics.
Angelina grew up loving the predictability and purity of equations in high school. As she remembers it: "For me, it was really just about the math, physics and the equations. I loved the predictability, the harmony, and the order when you can actually derive the answer that follows straight from the first principles." When the time came to choose a path, she went where the equations were densest: electrical engineering at the University of Belgrade. There she found a department called Physical Electronics, which she describes as focused on optical sciences, and nanoelectronics. "That seemed like such a niche and interesting field where you could apply all these interesting equations in physics. It was just magnetic for me," she says.
The light, she clarifies, wasn’t what hit her first:
Her PhD at Belgrade led to a postdoc in Thessaloniki, working in academia with the WinPhos research group. Then, in early 2020, what looked like a suspicious cold email turned out to be a career-defining opportunity.
"Celestial AI appeared out of nowhere. Our future CEO, David Lazovsky, reached out to professor Nikos Pleros that I was working with at the time. He (Dave, ed.) was interested in some of the research on optical computing that we did. And I thought to myself, this is too exciting to be skipped, what’s the worst thing that can happen?"
In retrospect this kind of moment always sounds simple. At the time, it wasn’t. Celestial AI was an early-stage startup in California with an ambitious, but unproven technology. Angelina would be working remotely from Greece, in a field where most serious work happens within a ten-minute drive of Santa Clara. As she puts it: “Applying to any job for a Silicon Valley company from Greece and expecting that you will be working remotely, from a different continent, sounded almost impossible.” Still, she proved it otherwise. It is worth pausing on what kind of decision this was. Angelina is a person grounded in mathematical precision, so the move to a commercially driven startup was a leap of faith. She had no way of knowing where it would lead her, but at the same time, her curiosity was too strong to be ignored.
By the time Angelina formally joined Celestial AI in October 2021, the team had already locked onto something the rest of the industry had not yet noticed. The dominant story in technology that year was compute. More processing power, larger models, faster training, and a whole industry chasing the same goal in slightly different ways. Celestial AI was looking at the layer underneath, at the increasingly urgent question of how the chips inside a data center actually talk to each other.
"The founders of Celestial AI and the early employees recognized that another problem will catch us blindsided, and that’s the connectivity issue. You can have a lot of GPUs, a lot of memory units, but if you cannot connect them fast enough, then all those resources are underutilized."
The thesis was pretty simple: as AI models grew larger, the problem would shift from how fast you could process data to how fast you could move it between processors and memory. The external validation came from somewhere unexpected.
"It was the first time that the industry turned and said, okay, the compute we might have at least temporarily solved, but the networking now is becoming a huge bottleneck."
Once model sizes exploded, the connectivity problem became impossible to ignore, and Celestial AI, which had been ahead of the curve for years, found Angelina in the middle of it.
The technology Celestial AI built, Photonic FabricTM, is best understood as a full-stack platform with a single mission. Photonics, analog and digital electronics, and the software layer, all plug-and-play by design. The mission: use light to connect chips at distances where copper used to do the job, so that AI can keep scaling without choking on its own wiring. The principle is to bring optical connectivity as close to the chip itself as physically possible. For years, the industry has been pushing optics ever closer, but it has always stopped at the chip's edge. The fibres dock at the perimeter, in what engineers call the beachfront, and like all beachfronts it is finite. A chip has only so much edge, and once that edge is full, there is nowhere left to add capacity. Photonic Fabric was designed to skip the beachfront entirely, bringing optical I/O into the interior of the package rather than relying solely on the chip's perimeter for connectivity. Angelina's explanation:
"If you could come optically underneath, then instead of having just a perimeter, you open the whole surface of the chip and the density of the connection that you can achieve optically, it grows by an order of magnitude."
The environment underneath a modern ASIC is genuinely hostile. Temperatures can swing by tens of degrees on very small-time scales, which leaves optics with two unappealing choices: be inherently insensitive to heat, or be controlled with the kind of precision that itself burns power and adds complexity. The industry's two standard answers both came with a catch. Ring resonators, the popular choice, were notoriously hard to stabilize thermally. Mach-Zehnder modulators, the conservative alternative, were too big to fit. Celestial AI reached for a third.
"We went with something that’s intermediate and kind of like a forgotten modulation technology in the interconnect world, and that is germanium silicon electro-absorption modulator," Angelina says. "And that one is compact enough, and it has a certain inherent level of temperature insensitivity, making its control much simpler than in case of rings. This choice allowed us to deliver unprecedented bandwidth density, up to one terabit per second per square millimeter."
In late 2025, the American semiconductor company Marvell announced it would be acquiring Celestial AI for up to $5.5 billion dollars, a deal that would close in February 2026. For Angelina, the news arrived the way these things usually arrive for the engineers actually building the technology.
"I learned about the acquisition the way that most employees actually learn about the acquisition: through the news," she says. "The first reaction was shock. Then, almost immediately, pride. This was the industry saying yes to what we had built."
If you map Angelina’s career on a timeline, the shape of it is striking. None of it was planned. The teaching role she took on as a PhD student gave way to a postdoc, the postdoc gave way to the call from Dave at Celestial AI, and Celestial AI eventually gave way to the acquisition that put her inside Marvell. What followed was less an absorption than a scale-up. Marvell was already offering connectivity at many levels of the data center, just not inside the chip, and Celestial AI walked in carrying the piece that was missing.
If there is a single thread running through Angelina’s view of her field, it might be impatience with overclaim. In March of 2026 she co-organised the OFC workshop: Chasing the Limit: On the Path to Photonic Scale-Up with Ultra-Low-Energy per Bit. The frustration that produced the title is one she has been carrying for a while: "What people want from optics is like a black hole of efficiency," she sighs.
It is a misconception, she argues. And it is mostly held by people outside the photonics community. Inside the field, the limits are well understood, and pretending they don’t exist damages the field’s credibility. Or as Angelina says:
The OFC workshop drew a standing-room crowd. What it produced was a working consensus. For photonics to compete seriously with electronics in scale-up domain, the target sits somewhere in the range of two to three picojoules per bit, somewhat below the current electronic state of the art, depending on reach and bitrate. Both numbers may move, but the gap is definitely worth chasing.
The same instinct that drew her to physics as a teenager is the instinct that 20 years later urges her to push back against people who want light to do impossible things. In doing so, she is helping reshape how future AI systems move data, one of the defining engineering challenges of the next generation of AI infrastructure. Light obeys the rules. The people working with it sometimes don’t.
At the end of the interview, when asked what governs how she works, Angelina offered two principles that read less like the rules of a physicist than like the lessons of someone who has built a life on the understanding that her career would never behave like an equation.
The first:
"You should start learning a tool when you actually have a use-case for it. I was pretty horrible at programming at the first years of my university. Everything seemed so abstract, but today, scripting is where my resting heart rate is lowest. Don't try to learn things just to check a box. I think that's what textbooks skip mentioning."
The second:
"Opportunities show up whenever in your life. Having a plan is great, but accepting an opportunity is even better."
These are the principles of someone who has accepted that life will not be derived from first principles. And here, finally, the tension resolves. Angelina trusts physics so completely that she can afford to be flexible everywhere else. The equations give her a baseline. From there, she leaps forward into the unknown. For a person who doesn’t like surprises, she has learned to make the most of them."
Dr Angelina Totović is a Principal Engineer, Photonics at Marvell , based in Thessaloniki, Greece. She holds a PhD in Nanoelectronics and Photonics from the University of Belgrade and has authored work that has been cited over 1,400 times in the scientific literature, with an h-index of 20. She holds multiple patents in photonic neural network architectures.
She is deeply embedded in the global photonics community. She has organised workshops at OFC, IPC, and Optica’s Advanced Photonics Congress, served as programme chair at the Integrated Photonics Research conference, and was guest editor of the IEEE/Optica Journal of Lightwave Technology Special Issue on Photonic Computing. She currently chairs the N2 Subcommittee for OFC 2027 and serves as Member-at-Large for IPC 2026. She speaks Serbian, English, and Greek.
Dr Angelina Totović
Principal Engineer at Marvell
- Articles