Power & Sites · United States

Data Centers, Gas, and Who Pays: Power Association Leaders Find Common Ground

Credited to Power Magazine · powermag.com · 7h ago

Photograph · Power Magazine

When the heads of three of the power sector’s biggest trade associations shared a stage at Experience POWER on Sept. 30, the most notable thing was how little they disagreed. Whether the question was data center backlash, grid security, or what to do about gas-plant emissions rules, the answer came back to the same point: The grid needs a lot more infrastructure, it will cost more than the old infrastructure did, and no regulatory shortcut will let the sector avoid either fact.

The session, “Inside the Issues,” was moderated by Mark W. Menezes, president and CEO of the United States Energy Association (USEA). It brought together Jim Matheson, CEO of the National Rural Electric Cooperative Association (NRECA); Tom Falcone, president of the Large Public Power Council (LPPC); and Rich Powell, CEO of the Clean Energy Buyers Association (CEBA). Co-ops, the largest public power systems, and the corporate buyers that include the hyperscalers all feel load growth differently. They still arrived at similar conclusions.

Data Centers Have a Social License Problem

Falcone said that growth projections at today’s level did not exist three years ago, and that he would not have predicted them. Some LPPC members, he said, “have to double or triple in size over the next 10 years.” He also said that 18 months ago he would not have expected people to be coming to LPPC members and saying, “You shouldn’t be building this.” The members’ answer is consistent: “Hey, we serve the needs. We aren’t the ones permitting the projects. We aren’t the ones deciding the land use.”

Powell, whose members include the hyperscalers, was the bluntest about the risk. He said the data center industry has responded with a White House ratepayer protection pledge and voluntary commitments on costs and water, yet opposition has become a dominant narrative in many communities. His worry, he said, is whether data centers “end up in the place that nuclear power plants ended up 40 years ago,” a reference to the long moratoria on new nuclear construction. The industry is weighing a more concerted national push to show that data centers benefit communities through tax bases, jobs, and other community benefits, he said. “The burden right now is on the data center community to prove that to communities in a more profound way than they were doing before.”

Powell also made the case against building private power plants next to data centers. His members would prefer to buy from utilities, he said. “We like grids. Grids are good.” He added that he sees large co-located projects as a sign the grid could not serve a load in time, citing a Meta facility in Ohio that, he said, faced a seven-year wait for interconnection.

Security Rules Need One Standard

On the recent executive order directing utilities to identify equipment that could pose security risks, particularly equipment from adversary countries, all three leaders accepted that the threat is real. “There are actors out there that want to infiltrate our system,” Matheson said, though he urged regulators to rely on “actual, truthful, fact-based information, and not fear and rhetoric.” Powell asked that any requirements not be over-broad or discriminatory toward particular technologies.

Falcone offered the most specific prescription. Utilities cannot see their suppliers’ suppliers, he said, and asking each supplier to disclose that to roughly 1,500 registered entities is unworkable. He pointed to the nuclear industry and Europe, which he said use certification of origin handled by a single body. He also raised aggregation: Individual inverter-based resources may be too small to pose a risk, but enough of them together might. Falcone cited a news report, which he said appeared about six months before the panel, of about a dozen inverter-based resources being opened up and unauthorized communications equipment being found in every one. A patchwork won’t work, he said: “It can’t be 1,500 entities on the bulk electric system, each trying to solve the problems themselves.”

Flexible Load: Welcome in Theory, Hard to Sell

All three praised the Federal Energy Regulatory Commission for not imposing a one-size-fits-all approach to large-load interconnection and for letting regional grid operators tailor their own rules. Falcone and Powell differed on how much the flexibility option will matter.

Falcone said one grid operator launched a product about six weeks before the panel that offers faster interconnection to flexible loads. He said that when LPPC members met with the operator, the question was whether anyone had signed up: “the answer is no.” A data center’s cost of interruption is so high, he said, that “you really don’t want flexible load. You want firm service.” He put it this way: “The amount of flexibility on the grid is determined by the buyer’s economics, not by the seller.” His conclusion: “There’s no magic bullet here.”

Powell did not dispute the economics but argued for a wider definition of flexibility. A buyer running high-value inference workloads may not curtail, he said, but could add energy storage or invest in virtual power plants that give the grid the same relief. He wants solutions “as performance-based, and technology and business-model neutral, as possible.”

Gas Rules and the Gas Bet

Matheson said the Environmental Protection Agency’s repeal of the Biden-era carbon capture requirement for existing coal and gas plants was a major win for cooperatives, who he said faced forced closures. But he warned that the rule’s efficiency requirement for new gas plants remains, and that NRECA’s outside consultants concluded it would limit combined-cycle plants to running only about 40% of the time.

Falcone argued the cost is the problem. A plant forced to run less means more plants. “It just makes the grid operate very inefficiently and raises costs without reducing carbon,” he said, adding that LPPC’s members span red and blue states. “We’re just for rules that work.”

Powell, who said he is a longtime carbon capture advocate, was candid about the technology’s status. “We do not in the United States today have a working carbon capture system up and running on a gas plant,” he said. Members such as Google have contracted for gas with capture, he said, and the first are scheduled to come online around 2030.

An audience question from a representative of Old Dominion Electric Cooperative on a possible gas overbuild brought out the shared worry. “Everybody’s betting on gas,” Matheson said, and “if you want to order a gas turbine, you can’t get one before 2032.” Falcone said fuel diversity protects the grid, but that as reliance on gas grows, redundancy, generally oil in a tank, costs money, and not every market has a revenue stream to pay for it. “The oil in the tank solves a lot of problems if you have a problem with a pipeline,” said Falcone. But for a for-profit generator, the logic is simple: “If you don’t pay me for it, I’m not going to do it,” he said. Powell questioned whether the country has an adequate framework for evaluating gas concentration risk, adding that the North American Electric Reliability Corporation has been calling for some kind of gas-electric coordination standard for years, along with reliability requirements for gas pipelines.

Permitting: From Briefings to Bill Text

The panel took place as a bipartisan Senate permitting package was expected to surface. Powell, citing briefings from the negotiators, described the outlines then circulating: faster environmental reviews, less litigation risk, protections against permits being pulled after construction starts, and long-haul transmission reform. Matheson said the consensus that Congress needs to act on permitting is stronger than at any point in his career. “Let’s take the win. Let’s move down the field. Let’s get some progress. We can live to fight another day,” he said, adding that he is a believer in taking “the partial loaf” when the full loaf isn’t available. Falcone was cautious: “Devil’s in the details, and we really need to see the bill text.”

The text arrived later that day. The Bipartisan American Affordability and Jobs Act was introduced by Sen. Shelley Moore Capito (R-W.Va.), chair of the Environment and Public Works Committee, and Sen. Mike Lee (R-Utah), chair of the Energy and Natural Resources Committee, along with ranking members Sheldon Whitehouse (D-R.I.) and Martin Heinrich (D-N.M.). According to the sponsors, the bill would let permitted projects keep their permits absent extraordinary circumstances, violations of law, or court order, and would require data centers to pay their transmission costs. The Bipartisan Policy Center (BPC), in an Oct. 2 explainer, described a 417-page bill that, among other provisions, would bar cost-shifting to other customers for new data centers of 20 MW or more. BPC said it expects a Senate cloture vote on the motion to proceed on Nov. 9 and that amendments are possible. CEBA issued a statement from Powell the same day calling the package “a generational opportunity to unleash abundant American energy.”

The session’s last exchange, prompted by a question about Pennsylvania’s infrastructure grade, brought the theme home. Falcone said LPPC members’ capital budgets have on average doubled in three years, which “means you need double the engineers and double the linemen and double the supply chain.” The sector will muddle through, he said, but the cost arrives either way. “Everybody has to pay more, and that’s an affordability challenge.”

Falcone then invited Powell to weigh in. “We are paying our way,” Powell said.

— Aaron Larson is executive editor of POWER.

Creating Value from Data: Maximizing Plant Performance Through Real-Time Analytics

Modern process industries are experiencing fluctuating market conditions and tight operational margins, leading chemical engineers to rely on real-time data to boost efficiency and reduce costs. Yet, many organizations are at different stages in their digital transformation journey. Some are just starting, while others are looking to optimize existing solutions. This webinar explores practical ways […]

Smart MRO: Using Technology and Assessment to Enhance Power Plant Performance

How One Electric Utility Is Using AI and Extreme Weather Modeling to Make Critical Grid Decisions

archives + podcasts + webinars + e-letters + video + exclusive content + events + much more!

For over 142 years, POWER has been the unwavering pillar of the power and energy industry, offering unparalleled insights on the latest trends, best practices, and news through our robust channels: print, digital media, and in-person events.

Power & SitesUnited States

DCD Studio: Workforce Development with Christian Goldsmith, Arcadis

Credited to Data Center Dynamics · datacenterdynamics.com7h ago

Useful

Tom Liddy speaks with Christian Goldsmith of Arcadis about tackling critical talent shortages in the data center sector through workforce planning, strategic training, and global external accreditation.

View the articleData Center Dynamics

Front Method

Power & Sites · United States

Optical illusion: Is co-packaged optics the savior the data center industry has been waiting for?

Credited to Data Center Dynamics · datacenterdynamics.com · 7h ago

Photograph · Data Center Dynamics

When Cisco took to the stage at the Optical Fiber Conference (OFC) 2023 to demonstrate co-packaged optics (CPO) on a 3.2TB switch, it is fair to say the technology did not receive a rapturous reception.

“I would say the industry response at that time ranged from indifference or a yawn to fairly negative,” Bill Gartner, SVP & GM, optical systems and optics at Cisco, tells DCD.

It wasn’t that three years ago the world didn’t care about the potential benefits of CPO. Rather, as Gartner explains it, the negative response stemmed from a number of concerns, one being that the technology would introduce vendor lock-in to an environment that traditionally allowed for a multi-vendor approach.

The cell tower climbers risking it all to help us stay connected

“There’s a switch, which makes up the bulk of the router and its costs, and then there are ports on the face plate into which you plug these pluggable optics,” Gartner explains. “But a customer has a choice of which pluggable optic they want to use, and there are many providers of pluggable optics. That gives them negotiating leverage and the ability to make sure that they've got supply chain diversity, which gives them a lot of benefit.”

In 2023, customers were fearful that CPO would force them into an all-Cisco or all Broadcom world, Gartner says, although he stresses that while Cisco is one of the aforementioned providers of pluggable optics, the company is not in the habit of forcing its customers to buy Cisco.

“Multi-vendor gives customers leverage, and customers like that leverage. It gives them peace and comfort that they're not going to be held hostage by a single source.”

While vendor lock-in remains an issue that CPO has yet to fully resolve three years later, it appears that the indifference Gartner witnessed on the OFC stage has vanished, replaced by an industry ready to shrug off all concerns and welcome the technology with open arms.

What is co-packaged optics?

Co-packaged optics emerged as a concept around the start of the decade, when the data center industry began discussing the need for a solution to the power and bandwidth bottlenecks it was increasingly experiencing.

“In the old days, your electrical connections were copper. Your data center networks were connected by Ethernet, and everything was good,” says Roy Chua, founder and industry analyst at AvidThink. “Then, as speeds increased, copper became unable to handle the volume of traffic, so the industry turned to fiber optics,” which Chua notes was also good, until it wasn’t quite good enough anymore.

Fiber optics requires the use of lasers to convert electrical signals into optics, or photons, so they can be transported over the glass fiber, then reconverted from optical signals back to electrical ones when they reach the other side. “So that's all wonderful, except that all takes a certain amount of energy,“ Chua says.

In today’s data center environments, most electrical to optical conversions take place via pluggable optical transceivers, known as digital signal processors (DSP), which connect to the switch. However, DSPs are expensive and power hungry, meaning that although they have clear benefits over copper interconnects, they are not quite the power-efficient option the industry is looking for.

Enter: Co-packaged optics. On a very basic level, unlike pluggable optics, where the transceivers are kept separate from the main chips, CPO integrates optical and electrical components, such as optical engines and switching ASICs, on a single packaged substrate.

“What co-packaged optics does is move the conversion onto the actual silicon itself. It moves those lasers onto the silicon, eliminating the need for any external units,” Chua explains. “And by putting those lasers inside the silicon, you are able to save power. By co-packaging these elements, it allows for more consolidation while lowering costs and energy use.”

However, while the concept might seem relatively straightforward, the practicalities are less so. “As you can imagine, manufacturing of co-packaged optics can be complicated because you have to manufacture everything – your wave guides and your modulators – all together.”

Therefore, it's unsurprising that another early concern regarding the technology was its potential cost, with Chua noting that “people felt it would not be sufficiently reliable and cost-effective enough to produce and take to market. “

The reliability of the technology has also come under scrutiny from customers, with questions arising as to how to go about replacing a laser that has failed when it's part of a monolithic structure, particularly given the so-called blast radius of optics, and the high failure rate of lasers.

The industry has responded by developing a hot-swappable laser source known as an External Laser Small Form-Factor Pluggable (or ELSFP) while keeping everything monolithically integrated. Gartner says that although this solution certainly helps to address the challenge, it’s something of a half measure that doesn’t succeed in solving the issue entirely.

The shifting optical landscape

According to AvidThink’s Data Center Networking in 2026 report, CPO is one of several new architectures that are driving a “structural shift” in the optical interconnect landscape, namely removing the need for DSPs from optical transceiver modules.

In addition to CPO, there’s what Chua describes as a half-step, known as linear pluggable optics (LPO), and a three-quarter step called near-packaged optics, with LPO currently the most popular of the three. There’s also eXtra-dense pluggable optics, or XPO, which was developed by Arista but described by Chua as less of a pure technology and more of a packaging form factor. [See Box Out].

While co-packaged optics is not a new concept, AvidThink’s report notes that the technology reached a milestone this year, transitioning from “perpetual promise to production deployment,” with both Chua and Gartner agreeing this has been largely driven by announcements made by Nvidia CEO Jensen Huang at GTC 2026.

At the March conference, Huang indicated that Nvidia’s Spectrum-X CPO switches would be in production in H2 2026, and officially announced that the chip giant’s 2028 Feynman architecture would introduce NVLink 8 CPO, bringing silicon photonics directly to the GPU-to-GPU scale-up interconnect for the first time.

AvidThink says this is architecturally significant as, until now, CPO has been applied to scale-out switch-to-switch connections. Applying it to the latency-critical scale-up domain signals that CPO technology has achieved the maturity and performance required for the most demanding interconnect tier.

“Until Jensen said CPO, no one cared about CPO. No one had any idea what CPO was, and now I get asked that question all the time,” Chua laughs. He also questions whether anyone in the industry would have been saying widespread adoption of the technology would take place in 2028 had Huang not gotten on stage and made the Feynman announcement.

“Without Jensen sort of stoking the fire, I think it would have been the 2027 - 2028 conversation, and Jensen basically made it the 2025 - 2026 conversation,” Chua says, adding that “to be frank, I wasn't expecting it.”

While Chua says Huang has put a rocket under the industry – with Broadcom in particular being incredibly aggressive around its timelines for deployment in the wake of Huang’s comments – the AvidThink report does highlight that there have been a number of other notable CPO announcements from the last 18 months.

These include AMD’s January 2025 acquisition of Enosemi; Lightmatter unveiling Passage L200, the world’s first 3D co-packaged optics product; Broadcom announcing its Tomahawk 6 CPO variant ‘Davisson’; Marvell’s acquisition of Celestial AI; Coherent demonstrating multi-technology CPO at OFC 2026; and Ayar Labs and Wiwynn announcing a joint reference design for optically connected, rack-scale AI systems.

Furthermore, in March 2026, Ciena, Coherent, Marvell, Molex, Samtec, and TeraHop announced the formation of the Open CPX MSA (Open Co-packaging Multi-Source Agreement) to develop the specifications for optical engines “required to enable a broad ecosystem of interoperable co-packaged and near-package interconnect solutions.”

Where are we now?

While all of these announcements appear to indicate significant demand, Gartner says it's important to note that a lot of work in the CPO space was initially vendor, not customer, driven, as demonstrated by those poor early responses.

“[In 2023] Cisco and Broadcom announced these CPO demos and didn't get a great customer response. Then Nvidia announced that they were going to deploy CPO and, quite honestly, stimulated a lot of excitement in the industry – I think maybe irrational exuberance is the way I would characterize it though.”

Although Cisco had already demoed a co-packaged optics prototype back in 2023, Gartner says the company has not yet announced a commercial CPO line, although he says that it does have CPO technology under development.

“We're playing this a bit cautiously,” he explains. “We don't want to come across to our customers like we're not going to give them a choice.”

Gartner says Cisco’s initial CPO offering will likely be in the 100-terabit switch capacity range, and the company will continue to give customers a choice between pluggables and an embedded option, be that CPO or NPO.

“Again, we have not announced a product yet; this is something that's under development, and we're really going to be cautious about launching it based on customer demand. We want to see the customers are really pulling for this, as opposed to it being a vendor push.”

In the meantime, the company is developing LPO options for its Silicon One portfolio, with the first commercially available 800G LPO modules for its Nexus 9000 series switches and 8000 series routers released in February 2026.

“I think the challenge that we have to acknowledge here is that the industry ecosystem is evolving around CPO, but it’s not mature. The ecosystem around silicon is highly mature – we know how to put a package and substrates around silicon, and we know how to do that around optics, and build a pluggable optics module, but we don’t have an ecosystem that’s mature and can put those two together,” Gartner says.

“There's a lot of invention right now around things like connectors for CPO, substrates that support the optics, and optical chiplets, but that's an immature ecosystem, and it really has to mature pretty significantly before we'll see this technology make sense at volume. Anybody can produce a hundred of these things, but when we have to talk about delivering thousands or millions, that ecosystem really needs to be mature, and that's going to take some time.”

Despite all the excitement, Chua also acknowledges that, at the moment, CPO is being deployed at “a very small scale.” However, he says all early signs are promising, and the data show that the technology will be able to scale up and be deployed into production.

In particular, Meta’s deployment of the Broadcom Tomahawk 5 ‘Bailly’ CPO switch in its data centers resulted in a 0.34 percent annual link failure rate during the first million hours of testing, making it roughly five times more reliable than traditional pluggable transceivers that were used as control groups.

As a result, Chua says he can’t see anything on the horizon that would block the mass adoption of co-packaged optics in the near future, and he expects to see some level of volume shipments from Broadcom of its Tomahawk 6 ‘Davisson’ CPO variant.

For Gartner, however, whether the reality of CPO lives up to the hype remains to be seen, especially with regard to the timelines being touted by vendors such as Nvidia and Broadcom. While shiny and new will be enough for some people to jump on board the co-packaged optics train, for him, the real turning point for the technology will be when all the other technologies currently on offer hit their ceiling, making CPO a necessity.

“Today, the SerDers rate coming off the silicon is 200G per line, and at that rate, in my view, it's a toss-up as to whether CPO or LPO makes sense,” he says. “Once we get to 400G per line, then LPO becomes much more challenging... that is where the ability for LPO to continue serving the market diminishes, and the need for CPO becomes more real.”

Co-packaged optics is not the approach currently making waves in the optical interconnect landscape. Linear-pluggable optics (LPO), near-packaged optics (NPO), and also eXtra-dense pluggable optics (XPO) have all become part of the network interconnect conversation. But what are the differences between each offering?

Linear-pluggable optics work by replacing the digital signal processing (DSP) chip from the optical module and replacing it with high-linearity analog components. However, unlike co-packaged optics, LPO has a pluggable form factor, meaning that while it has a lower power consumption than traditional optics, it’s not as efficient as CPO.

Viewed as a three-quarter step on the way to co-packaged optics, near-packaged optics is an approach that sees the optical engine sit on the host board very close to the processor, whilst still maintaining its own separate, independent packaging. While highly efficient, it still loses a small amount of power due to the need for a short copper trace on the printed circuit board.

Touted as the most efficient optical interconnect approach, co-packaged optics integrates all the optical and electrical components directly onto a single substrate. However, the reliability of the technology has been called into question, and there remain some concerns about the maintenance of CPO given its monolithic architecture.

And finally, there’s XPO. Introduced by Arista Networks and supported by an industry Multi-Source Agreement that includes LightMatter, the high-density, liquid-cooled optical transceiver form factor has been designed specifically for large-scale AI and machine learning clusters. It’s claimed that when compared to traditional pluggable transceivers, this approach enables a 4x increase in switch rack density.