Models & Labs · Europe

British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted

Credited to The Register · theregister.com · 1h ago

Photograph · The Register

Revolut says that in less than a year it's managed to slash its AI coding inference costs by roughly 90 percent by making self-hosted models its default instead of relying primarily on commercial services like Claude.

Speaking at the AI Engineer conference in New York last month, David Abitbol, ops and infrastructure lead at the British fintech, described how it saw a 60 percent improvement in coding time and a 70 percent improvement in lead time following its rollout of AI coding tools to roughly 2,000 engineers. Revolut initially adopted GLM-5.1 before moving to GLM-5.3, which it now serves using 56 Nvidia B200 GPUs.

Prior to its switch to GLM, Revolut had been experimenting with AI, deploying Gemini in early 2025 and Claude Code in early 2026. "We were highly skeptical about the cost of AI. We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI, and so when we launched it we launched it with strict budget limits," Abitbol said.

Because of this, engineers only had access to a limited number of tokens, so the company began investigating self-hosting iyd own models. At $0.30 per million input/output tokens, self hosting GLM 5.1 cost about a tenth of using OpenAI's Codex or Anthropic's Claude Code. And, as it turned out, when tokens are cheap, you stop worrying about how many they use or what they use them for.

While Revolut did see improvements to coding speed, Abitbol noted the biggest bottleneck wasn't actually the code. Instead, he says the biggest gains came from letting the AI models deal with things like debugging and Jira and Confluence tickets. "When you have unlimited tokens, you're free to do that," he said. "You no longer have to save your tokens for the end of the quarter or the most difficult task."

Developed by Chinese open weights model dev Z.ai, GLM 5.1 isn't exactly a small model. At roughly 754 billion total parameters (40 billion activated for each token generated), the model requires roughly 380 GB of GPU memory just to hold the weights and another 6-26 GB per user to maintain context depending on precision.

To serve GLM-5.3 at scale, Revolut deployed 56 Nvidia B200 GPUs, equivalent in GPU count to seven eight-GPU DGX B200 systems. But as Abitbol noted, the greatest performance gains didn't come from adding GPUs, but rather tuning the model's hyperparameters along with switching to NVFP4 quantization, implementing disaggregated prefill and decode, and tuning context lengths.

One key takeaway was that supporting a one-million-token context window doesn't necessarily make such a large window useful or efficient. "Long context is a scam. You really need to be mindful of that," Abitbol said. "Even though you can serve one million tokens, 256 [thousand] is really what you need."

With these optimizations, Abitbol says that the company was able to achieve sub 3-second time to first token on context windows of around 256,000 tokens, and generation speeds of around 200 tokens a second, both of which he notes are the sweet spot for the company's software engineers.

"What you want to aim for is a sub-3 second time to first token. This is empirical evidence. As soon as you exceed 5 seconds, my Slack inbox starts to blow up," Abitbol said. "In terms of throughput, it doesn't really matter as much. We serve, in terms of interactivity, 200 tokens per second. People are fine [with that]. You don't need to go higher. The way people work has changed. We no longer look at the stream of thinking tokens. People run tasks in parallel and check in when the task is done."

And if throughput is a concern, he notes engineers who sign in after 5 pm are getting closer to 800 tokens a second.

However, it's worth noting that GLM hasn't replaced GPT or Claude entirely. The Chinese model is text only, so it isn't ideal for things like UI work. So instead, Revolut used a model router to direct vision queries to models like Gemini or GPT Luna while keeping code generation running on GLM.

"Open source is now a priority at Revolut, and our economics are favorable. The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception," Abitbol said.

Having said that, Abitbol acknowledges that Chinese open weights models may be a concern for privacy and security reasons, but emphasizes that the major concern there is accessing them via APIs running out of China. When run on-prem or on regional cloud resources, there's really nothing to worry about. ®

Power & SitesAsia-Pacific

Naver backs Brookfield's $9bn GPU deal with a ₩18.7tn guarantee

Credited to Rackline · Priya Nair · dart.fss.or.kr2h ago

Useful

Nvidia will pay ₩1.34tn on Nov. 6 for 6,558,435 new Naver shares, Naver said in a Korean filing on Nov. 3, after the won rate for the $1bn investment was fixed at ₩1,341.20 to the dollar. The placement depended on Naver first securing at least $9bn of computing infrastructure, which it did on Oct. 27 through a deal with Brookfield.

That deal is meant to ease the upfront cost of Naver's AI factory. The filings show who carries the risk. Brookfield puts up the money, Naver guarantees the rent, and the guarantee is far larger than the cost of the hardware.

Under the Oct. 27 agreement, a company to be set up by Brookfield's AI infrastructure fund, called Asset Co. for now, will spend up to $9bn by 2028 on GPUs and data centers and rent the computing to Naver Cloud. Naver plans to start at 55MW in the first half of 2027, reach 100MW by the end of that year and 200MW in 2028, with its GAK Sejong data center as the first hub.

Nvidia sits on both sides of the arrangement. Naver's plan names it as the GPU supplier, and the two companies will work together on Nvidia's next-generation DSX data-center platform. Its share purchase could not close until Naver had lined up the $9bn, so the equity and the rental deal arrive as one package.

Naver guaranteed Naver Cloud's payments under those contracts. Its filing puts the guarantee at ₩18.74tn, equal to 64.73% of Naver's equity, running to the end of 2034. The figure is the nominal total of service fees Naver Cloud would owe if the full $9bn is spent. Damages and penalties are covered too, but the filing says it cannot yet put a number on them. With an older ₩0.75tn guarantee, Naver's total guarantees now stand at ₩19.49tn.

At the rate fixed on Nov. 3, $9bn is about ₩12.1tn*. The fees Naver has guaranteed come to about 1.55 times* the cost of the equipment. The filings do not split that gap between financing, return and operating charges. Payments run for six years on each batch of chips from the day that batch is delivered.

Naver has guaranteed more than the kit costs

₩tn; Naver Cloud at end-2025, Naver equity as stated on Oct. 27, 2026

₩tn

View chart data

Naver has guaranteed more than the kit costs
Item	₩tn
Naver Cloud equity	0.66
Naver Cloud assets	2.33
$9bn of equipment*	12.07
Naver's guarantee	18.74
Naver equity	28.95

Source: Naver filings to Korea's DART, Oct. 27 and Nov. 3, 2026; Rackline estimate for equipment*Rackline estimate: $9bn converted at ₩1,341.20 per dollar, the rate fixed on Nov. 3.

The unit that owes the money is small next to that bill. Naver Cloud had ₩1.55tn of revenue in 2025, ₩165.6bn of net income and ₩656.9bn of equity. The guarantee is about 12 times* its annual sales. Brookfield's real counterparty is the parent, which will charge Naver Cloud 0.4375% a year for the guarantee, worth ₩670.7bn over the term on the same full-spending assumption.

Nor does the structure take the assets off Naver's books. Brookfield's fund will own all of Asset Co.'s capital, but Naver says it will control the company under Korean accounting rules and consolidate it as a subsidiary. Asset Co.'s equipment and obligations would then sit in Naver's group accounts. Naver's release says the guarantee itself costs nothing and moves no cash. That is true only for as long as Naver Cloud pays.

What Naver gains is timing. It says usage fees will be tied to the AI factory's revenue to ease the early cash strain, and chips will be bought in batches as demand appears. Nvidia's money adds equity. Naver plans to spend ₩4.56bn of it this year, ₩626.7bn in 2027 and ₩709.9bn from 2028 on the AI factory business.

The stronger won trimmed the proceeds. In July, Naver had estimated the placement at 7,241,564 shares, or ₩1.48tn at ₩1,480.9 to the dollar. The final issue is 683,129 shares* smaller and brings in ₩139.7bn* less. The new shares equal about 4.1%* of Naver's enlarged share count, list on Nov. 27 and are locked up for a year.

Who does what in Naver's AI factory

Table of the Naver AI factory deal: Brookfield's Asset Co. buys up to $9bn of equipment, Naver Cloud rents it, Naver guarantees ₩18.74tn, Nvidia buys $1bn of shares.
Party	Role	Filed terms
Brookfield fund, via Asset Co.	Buys and owns GPUs and data centers	Up to $9bn by 2028; to be consolidated by Naver
Naver Cloud	Runs the AI factory and rents the computing	Fees run six years from each batch's delivery
Naver	Guarantees Naver Cloud's payments	₩18.74tn to end-2034; 0.4375% annual fee
Nvidia	Buys new Naver shares	6,558,435 shares, ₩1.34tn; paid Nov. 6
Capacity	Build-out targets	55MW H1 2027; 100MW end-2027; 200MW 2028

Source: Naver filings to Korea's DART, Oct. 27, 2026 and Nov. 3, 2026; Naver release, Oct. 27, 2026

For Brookfield, this is Naver credit with chips attached, not a wager on what used GPUs will fetch. For Naver shareholders the test is how fast the 200MW fills. If customers arrive on schedule, the revenue-linked fees work as designed. If they do not, the guarantee still covers six years of payments on every batch Naver Cloud has taken.

* Rackline estimate, calculated from Naver's DART filings of Oct. 27 and Nov. 3, 2026: $9bn at ₩1,341.20 per dollar (the filing does not state the rate behind the ₩18.74tn); ₩18.74tn divided by that sum and by Naver Cloud's 2025 revenue of ₩1,554.5bn; July versus Nov. 3 share and won amounts; 6,558,435 new shares over 158,684,979 after the issue.

Source · dart.fss.or.kr

Front

Power & Sites · Asia-Pacific

Samsung SDS breaks ground on Gumi AI Data Center in South Korea

Credited to Data Center Dynamics · datacenterdynamics.com · 4h ago

Photograph · Data Center Dynamics

Samsung SDS has broken ground on a new AI data center in Gongdan-dong, Gumi-Si, South Korea.

Dubbed the Gumi AI Data Center, the facility will be developed in phases, starting with 60MW of capacity and ultimately scaling to 120MW.

The data center will be housed on a 498,165 sq ft (46,281 sqm) site. With construction now underway, the data center is expected to be completed in late 2028.

The data center project is part of Samsung's commitment to invest 60 trillion won ($44.671bn) in establishing the Yeongnam region as a "Global Physical AI Innovation Cluster." Samsung SDS announced last January that it would invest 427.3 billion ($318 million) in an AI data center in Gumi, and signed a memorandum of understanding (MoU) with North Gyeongsang Province and Gumi City for the construction of the project.

The company is also investing in manufacturing and robot automation, all-solid-state batteries, package substrates for AI servers, and MLCCs (Multilayer Ceramic Capacitors).

The data center will have a hybrid cooling system, combining air cooling in general server areas for networking and storage, and liquid cooling for the high-density GPU server areas.

The data center will host in-house GPUs that can offer compute to customers, but will also have a colocation business to host other companies' hardware.

Lee Jun-hee, president and CEO of Samsung SDS, said: "AI is transforming the way companies work and the competitiveness of industries. For these changes to translate into tangible results, a robust foundation capable of processing vast amounts of data and operating AI stably is required, and the Gumi AI Data Center will fulfill that role."

He added: "Samsung SDS will combine its accumulated experience in data center construction and operation with its cloud and AI technology capabilities to create an AI infrastructure that customers can trust and utilize."

Samsung SDS is the IT and cloud unit of Samsung. In August 2026, it revealed it would be launching an NPU-as-a-Service (NPUaaS) offering in South Korea with FuriosaAI's RNGD NPUs.

Samsung SDS is deploying the RNGD accelerators at its data center in Dongtan, which opened in March 2023. Named Dongtan IDC, the data center spans six floors with four server rooms, and has a PUE of 1.12 enabled by its use of natural air cooling.

Samsung SDS claims to operate 18 data centers from which it offers cloud services, and with the Gumi site, it will have five in South Korea.