Models & Labs · United States

Anthropic is working on a system to prove AI models’ identity

Credited to TechCrunch · techcrunch.com · 2h ago

Anthropic is working on an open-source system to verify which AI model produced a given output, according to a blog post published Tuesday. The system, called FlashProve, is built to provide what researchers call “provable inference” – a cryptographic proof, essentially a mathematical guarantee, that a given output came from a specific model. It’s meant to let users confirm that the answer they received really came from the model they think they’re using.

Similar to cryptographic email signatures that let recipients confirm a message really came from its claimed sender, FlashProve would be robust against impostors attempting to disguise one model as another, even when the impostor is the lab that built and runs the model. In particular, the researchers note that the system could be used to prevent an AI company from secretly swapping in a cheaper model to cut costs, or a more expensive model to inflate its performance results.

FlashProve could also be used to verify claims by third parties that a specific task was done by an advanced model. “A developer might claim that their software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company might claim that its business plan was given the go-ahead by an appropriate model,” the researchers write. “How can a user verify these claims?”

To work, the system requires AI labs to publish a “fingerprint” of each model they release that can later be matched to specific outputs and verified cryptographically. Crucially, the fingerprints don’t reveal a model’s weights, the internal numbers that define how it behaves and which closed-model makers like Anthropic and OpenAI keep private. That means labs could publish the fingerprints without giving up their most valuable asset, enabling the system to work with commercial models, too. The system is also designed to be lightweight, so verification could be performed on a local device.

As part of the announcement, researchers released a detailed technical specification and working code for various elements of the system, which is publicly available under an open-source license on GitHub. Research in the system is ongoing, and there are no plans to deploy it at Anthropic or any other lab.

“This research post is meant to share intermediate progress, and not a completed research agenda,” researchers explain in the announcement post. “We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it.”

Power & SitesUnited States

Elon Is Doing It Again: SpaceX Will Be The World’s Biggest AI Cloud

Credited to SemiAnalysis$ · newsletter.semianalysis.com4h ago

Useful

Making 13GW in 1.5 Years A Reality, Repurposing Rocket Ship Campuses, A Datacenter at TeraFab?

View the article · $SemiAnalysis

Front

Power & Sites · United States

Marvell Cuts To The Front Of The Datacenter Optics Line With Lumentum Deal

Credited to The Next Platform · nextplatform.com · 3h ago

Photograph · The Next Platform

For a long while, like several decades now, many of us have been waiting for silicon photonics to come to the datacenter en force to break the tyranny of distance and latency between compute, memory, and storage.

With light pipes that have enough bandwidth and low enough latency, we could stop building discrete and nearly identical computers and instead take banks of components and configure them on the fly into distinct architectures of computers whose feeds and speeds more perfectly matched whatever mixed workloads were running on systems and clusters.

It was a nice idea, but it never quite came to pass. It was just too expensive and too risky to not push electrical signaling to the limits of physics. And so the introduction of photonics alternatives was delayed, and high volume manufacturing investments for photonics suffered from the classical chicken-egg problem.

But then along came GenAI and its insane high bandwidth and low latency requirements. Something that classical machine learning did not require to the same degree, and within a matter of years all of those pie-in-the-sky startups that we have talked to over the decades – many with founders and techies with deep heritages in optics from the telecom and service provider industry – started to look not so crazy after all. And here we are, standing on the precipice of mass commercialization of photonics within systems, not just being used in the optical transceivers connecting systems within datacenters and regions and now even connecting regions into global compute farms. And it is all happening because there appears to be no limit, as yet, for what the hyperscalers, cloud builders, and AI model builders will pay for the most advanced systems hardware that Earth can muster.

Composability is not the explicit goal, but breaking through the “copper wall,” as many in the industry are calling it, certainly is. And thankfully, we may get the composability that I have been advocating for as a side benefit, because at some point, you can just put together trays or racks of optically connected CPUs, GPUs, XPUs, DPUs, and memories, interconnect them all with photonics and configure ratios of them as you see fit as workloads run.

That is why chip maker and sometime chip shepherd Marvell is acquiring photonics expert Lumentum for $80 billion in stock and assuming its debts to yield and enterprise value of $115 billion. This is nearly twice as much as the $67 billion that Dell spent to acquire VMware and EMC as a sorta matched set in October 2015 and it is four times what Avago Technology spent in May of that same year to buy chip maker Broadcom, laying the foundation of the chip and software behemoth (which has owned VMware for several years now) that is also a chip shepherd for the hyperscalers, cloud builders, and AI model builders, competing with Marvell. Broadcom spent $61 billion in May 2022 to acquire VMware, of course, to get its enterprise platform, which also includes Brocade Fibre Channel switches, and systems software from legacy suppliers CA Technologies and Symantec.

Money was relatively cheap back then, and it is a good thing that stock funny money is cheap now given that real money is considerably more expensive back then. Marvell’s prospected in the GenAI era have been improving fast, and its ability to acquire Lumentum, one of two key suppliers of the indium phosphide lasers that enable the use of optical interconnects for scale up networks across racks and maybe even rows and scale in networks that link elements of a compute engine to the outside world. The other one is, of course, Coherent, which like Lumentum, received $2 billion to expand its manufacturing back in March of this year. Marvell itself got a $2 billion infusion from Nvidia a few weeks later to help its chip shepherding customers adopt NVLink ports on their custom XPUs and maybe some CPUs, too.

If Nvidia wanted to truly control supply for optical components, it could buy any of these companies, or all three of them, with stock funny money, given its roughly $5 trillion market capitalization. But that would raise the eyebrows of the antitrust authorities. And buying just Lumentum or Coherent alone would only give Nvidia around half of the supply of lasers (and possibly optical circuit switches) that future Big Green AI systems will employ. Clearly, it is better, then, for there to be multiple suppliers to increase supply and, in theory, grind against each other to try to lower costs. This is not really working in the HBM market that Nvidia and all GPU/XPU suppliers are dependent upon. No one, not even Nvidia, can get the HBM capacity it wants at any price. Which is the gating factor on the entire GenAI industry. And optical interposers and lasers are definitely, without question, going to be another gating factor in the coming years.

This is the important thing about this Lumentum deal.

Nvidia can’t buy one vendor because neither Lumentum nor Coherent will individually make enough lasers to supply Nvidia’s needs. Nvidia can’t buy both of them because of antitrust issues, and it can’t start up its own operations in a timely fashion, either. But Marvell, which has a fledgling and fast-growing custom CPU and XPU business with Amazon Web Services, Microsoft, and Google, absolutely can but either Lumentum or Coherent and thereby guarantee that in the coming years, it can tilt laser supply towards its custom CPU and XPU business and still make money selling to the likes of Nvidia, AMD, Intel, hyperscalers, cloud builders, and even AI model builders creating their own chips with the help of chip shepherds like Broadcom, Marvell, Alchip, and MediaTek.

So while Marvell chief executive officer Matt Murphy can talk about this deal being about becoming “a category leader for the next generation of AI connectivity,” I think it is more about seeing that the order book for photonics lasers is already sold out through early 2029 and that this is going to be a gating factor in that Marvell custom chip business.

It is like Nvidia being rich enough to buy Micron – which it was not when the GenAI boom started, or it might have thought about it. Marvell, with a market cap of around $150 billion when the deal was announced, can spend $80 billion of that to effectively merge with Lumentum, which will have three seats on the Marvell board, and assume another $35 billion in debts and obligations.

The opportunities for both Marvell and Lumentum are both large, which you can see in the chart below as they are both starting to hockey stick revenues and profits, but this deal is also relatively large compared to the revenue and profit streams now. As was the case with Marvell’s acquisition of PHY maker Inphi several years ago, and ditto for the very expensive $3.25 billion acquisition of Celestial AI in December 2025 for its photonics interposer technology. (This changed Marvell’s story, and also positioned the company for the optical on-package compute engine future that its custom chip customers want it to de-risk.)

I have aligned the quarters for the companies, which are one month out of phase. Lumentum had a loss of $7.16 billion two quarters ago, as you see from the slopes of that orange dashed trough, which was related to converting debt to stock. Aside from the $7.76 debt being paid off this way, operating income was $279.3 million for the company against $1.01 billion in sales; it is hard to say what net income might have been, but call it $250 million or so given in its most recent quarter, Lumentum had $360.3 million in profits against $1.26 billion in sales. By the way, Lumentum is more profitable than Marvell, which had $308 million in net income against $2.74 million in sales in its most recent quarter. But that profit level for Lumentum is brand spanking new, coinciding with the rise of lasers in future systems coming down the pike from Nvidia and others.

Murphy thinks this deal for Lumentum is better than the Inphi deal was, which by the way has been the dominant driver of its AI-related business, not custom chip shepherding sold to the likes of AWS, Google, and Microsoft.

“We are at the inception of the transition as the copper wall moves from scale-out to the next wave, which is going to be scale-up and then scale-in beyond that,” Murphy explained on the call explaining the deal to Wall Street. “And it's hard to see sometimes. If you look back at Inphi, when we acquired it, the company looked like we were paying a very high multiple for it. It was sort of not clear how big it could really be. But you can see when you capture a market inflection at the right time with the best asset and the best team and you execute well, you can really outperform.”

Muphy added that the scale of the Lumentum business over the next few years is about an order of magnitude larger than what it expected out of Inphi when it bought it five years ago – and the orders are on the books for Lumentum, it is not just a projection.

What Marvell also wants to do by acquiring Lumentum is break into the bigtime in the semiconductor industry, to bypass MediaTek and Qualcomm and to start nipping at the heels of Intel and maybe AMD:

Isn’t that a fun chart? Marvell is crowing about how it will be the largest pure-play provider of datacenter connectivity chips, from the compute engine all the way up to across datacenter regions, but that is a bit silly. First, it does work on compute engines, just like Nvidia, Broadcom, AMD, and Intel, and second, Nvidia and Broadcom will have huge datacenter connectivity businesses that are several times larger than all of the expanded Marvell + Lumentum (shown as NewCo in the chart above).

The way Murphy & Co see it, Marvell had a total addressable market of $400 billion in calendar 2030, driven largely by its scale out and scale across chips for AI clusters but also including some scale up and scale in revenues inside nodescale and rackscale systems that make up these clusters. Lumentum had a $90 billion TAM before the deal. While there was talk of expanding the TAM, supply chains are limiting the TAMs, and as it turns out there is an additional $60 billion that Marvell thinks it can chase – mostly through near-package optics (NPO) and co-packaged optics (CPO) chips and packaging. This may be an undercounting, especially if custom chip designers want someone who can help with the entire scope of fabrics in their AI system designs.

This seems very likely. What is driving this need for full stack optical fabrics it is the need for larger compute and storage complexes that pack a ton more stuff in a lot more surface area, turning the row into the new rack and the system board into the new compute engine. This cannot be done with copper signaling:

Muphy explained that for every layer in the network hierarchy that you drill down into, from connecting distributed datacenters (scale across) to connecting rackscale systems within a datacenter (scale out) to connecting elements in a rackscale system (scale up, with coherent memory across the elements) to connecting compute and storage components on a package or across a tray within that rackscale system (scale in), you need another 10X increase in bandwidth.

The reverse is also true. As you go from the package or tray out to spanning the globe, there is no way to get enough electrical or optical signaling lanes out of devices to keep the bandwidth constant as you climb the network hierarchy. The best you can do is cut it by an order of magnitude to extend the reach of the link and wait for data to be exchanged across successive and nested layers of the hierarchical fabric.

Here is how the Marvell and Lumentum products today address the scale across and scale out networking, as well as how on-package NPO circuits can be used for giving scale up coherent networking to switches and XPU accelerators:

This is the scale up and scale in future that Marvell is chasing and that requires deeper integration of fabrics with accelerators and traditional electronic switches to drive up bandwidth while driving down latency and power consumption per bit per second:

All of this requires experience in getting chips designed and out the door on time, and both Marvell and Lumentum have delivered billions of chips in their respective histories. Whoever breaks optics into the datacenter bigtime for scale up and scale in will also have to have deep expertise in indium phosphide (for lasers), silicon germanium (for heterojunction bipolar transistors, which are used in the analog front ends of optical links), and gallium arsenide (for another kind of laser).

Marvell and Lumentum together have such expertise. Adding them under one roof makes it easier for customers to get comfortable about vertical integration and some kind of volume leverage, but you also have the “all my eggs are in the same basket” problem.

As we are found of saying in our household raising children: You can’t quit, and you often can’t win, but you can break even.

There is one more thing to consider in why this is an interesting deal for Marvell. Don’t count out those optical circuit switches as a replacement for Ethernet switches in AI back end networks, which is in the scale up domain, sorta sometimes maybe.

Google first used its “Palomar” optical circuit switch, which is based on reconfigurable arrays of micromirrors to create a backend network lashing together nearly 10,000 of its TPU accelerators together in reconfigurable topologies, back in 2020 with the TPUv4 accelerators. (The homegrown OCS is sometimes also known by the “Apollo” codename used for the entire OCS project, but Apollo really refers to the architecture, not the particular MEMS switch. Even Google calls it Apollo sometimes when it means to say Palomar.)

If Google can reduce power consumption and latency with OCSes, then it stands to reason that other hyperscalers, cloud builders, and AI model builders will want to do the same. And if not, Marvell has Innovium TeraLynx Ethernet switch ASICs they can slide into the backend.

One last thing: This deal is going to take forever to close, with an expected completion sometime in the first half of 2028. That is a million years in AI Land. And it is hard to see how the Chinese antitrust regulators are not going to mess with it. Europe might get involved. How Europe or China have any say at all is an interesting thing to contemplate. In any event, 20 months from now is a long time for this deal to get done. Someone could make counteroffers in that time, which would be interesting but also disruptive. This deal also immediately puts Coherent into play, too. Just like Intel acquiring FPGA maker Altera a decade ago put Xilinx immediately into play.