OCP 2025 Meta keynote: Scaling the AI Infrastructure to Data Center Regions

Posted on by Alan Weissberger

At the OCP Global Summit 2025 in San Jose, CA, Meta detailed its strategy for scaling AI infrastructure to regional data center deployments, emphasizing open, collaborative, and highly scalable designs to support growing AI workloads. The October 14th keynote presentation by Meta’s VP of Data Center Infrastructure, Dan Rabinovitsj, discussed strategies for deploying and operating AI at scale across various data center regions at OCP 2025. The session highlighted innovations for building AI-ready data centers, focusing on open hardware, power innovation, and challenges in next-generation AI infrastructure.

Initiatives discussed included: new Ethernet standards for AI clusters, integration of the Ultra Ethernet Consortium standard, Meta’s vision for open networking hardware, AMD’s “Helios” rack-scale AI platform, MSI’s integrated OCP solutions, next-gen liquid cooling, and solutions for distributed and edge AI.

Rabinovitsj highlighted Meta’s contributions to open standards and hardware innovations, including the Open Rack Wide standard and advanced networking concepts for AI clusters.

Meta also announced several new milestones for data center networking:

  • The evolution of Disaggregated Scheduled Fabric (DSF) to support scale-out interconnect for large AI clusters that span entire data center buildings.
  • A new Non-Scheduled Fabric (NSF) architecture based entirely on shallow-buffer, disaggregated Ethernet switches that will support our largest AI clusters like Prometheus.
  • The addition of Minipack3N, based on NVIDIA’s Ethernet Spectrum-4 ASIC, to our portfolio of 51 Tbps OCP switches that use OCP’s SAI and Meta’s FBOSS software stack.
  • The launch of the Ethernet for Scale-Up Networking (ESUN) initiative, focused on making Ethernet suitable for connecting high-performance processors, or GPUs, within a single rack by emphasizing requirements like low latency, high bandwidth, and lossless transfers. Meta has been working with other large-scale data center operators and leading Ethernet vendors to advance using Ethernet for scale-up networking (specifically the high-performance interconnects required for next-generation AI accelerator architectures.

OCP Summit 2025: The Open Future of Networking Hardware for AI

Key hardware projects discussed by Meta included:
  • Open Rack Wide (ORW) standard: Meta introduced the ORW specification, a new open standard for double-wide equipment racks designed to meet the extreme power, cooling, and serviceability demands of next-generation AI systems. AMD, a partner of Meta, showcased its “Helios” rack-scale platform built to be compliant with this new standard.
  • Networking fabrics for AI clusters: Meta detailed its networking architecture, revealing the following innovations:
    • Disaggregated Scheduled Fabric (DSF): An updated version of DSF was discussed (see below), which now provides non-blocking interconnects for clusters of up to 18,432 XPUs (AI processors), enabling communication between a larger number of GPUs.  
    • Non-Scheduled Fabric (NSF): Meta unveiled NSF, a new fabric for its largest AI clusters, which runs on shallow-buffer, disaggregated Ethernet switches to reduce latency. NSF is planned for Meta’s upcoming multi-gigawatt “Prometheus” clusters. See next section below for details.
  • FBNIC: Meta announced FBNIC, a network ASIC of their own design.
  • 51T switches: Meta revealed new 51T network switches, which utilize Broadcom and Cisco ASICs.
  • Next-generation optical connections: For faster and higher-capacity optical interconnections, Meta discussed its adoption of 2x400G FR4-LITE and 400G/2x400G DR4 optics for its 400G and 800G connectivity.
  • Sustainable hardware: As part of its 2030 net-zero goals, Meta presented a new AI-powered methodology for tracking and estimating the carbon emissions of its IT hardware. The methodology will be open-sourced for the wider industry

……………………………………………………………………………………………………………………………………………………………………………………………………………………………………

Deep Dive into DSF and NSF:

1. Disaggregated Scheduled Fabric (DSF):
DSF is designed to provide a highly efficient, lossless, and scalable network. First introduced at OCP in 2024, Meta announced a major upgrade to its design. 
  • Non-blocking scale: An updated, two-stage architecture for DSF can now support a non-blocking fabric for up to 18,432 XPUs (AI processors). This allows all-to-all communication between a significantly larger number of GPUs without performance degradation.
  • Proactive congestion avoidance: DSF uses a Virtual Output Queue (VOQ)-based system to manage traffic flow. By scheduling traffic between endpoints, it proactively avoids congestion before it occurs, which improves bandwidth delivery and overall network efficiency.
  • Open and standardized: The fabric is built on open standards like the OCP-SAI (Switch Abstraction Interface) and is managed by Meta’s own network operating system, FBOSS. This vendor-agnostic approach allows Meta to use components from different suppliers and avoid vendor lock-in.
  • Optimal load balancing: Traffic is “sprayed” across all available links and switches, ensuring an equal load and smooth performance for bandwidth-intensive workloads like AI training. 
2. Non-Scheduled Fabric (NSF):
Meta unveiled NSF as a new fabric specifically for its most massive AI installations, including the multi-gigawatt “Prometheus” cluster scheduled for 2026.
  • Low latency: Unlike DSF, which relies on scheduling, NSF operates on shallow-buffer, disaggregated Ethernet switches. This reduces round-trip latency, making it ideal for the most latency-sensitive AI workloads.
  • Adaptive routing: The NSF architecture is a three-tier fabric that supports adaptive routing for effective load-balancing. This helps minimize congestion and ensure optimal utilization of GPUs, which is critical for maximizing performance in Meta’s largest AI factories.
  • Disaggregated design: Like DSF, NSF is built on a disaggregated design. This allows Meta to scale its network by using interchangeable, industry-standard components instead of a single vendor’s closed system.
3. A dual-fabric strategy for the future:
Meta’s decision to pursue both DSF and NSF reflects its strategy for tackling the diverse and growing networking challenges posed by modern AI.
  • DSF: Provides a high-efficiency, highly scalable network for its large, but still modular, AI clusters.
  • NSF: Is optimized for the extreme demands of its largest, gigawatt-scale “AI factories” like Prometheus, where low latency and robust adaptive routing are paramount. 
This parallel, dual-fabric strategy allows Meta to build and operate AI infrastructure with unprecedented scale, performance, and flexibility, using open standards to accelerate innovation and reduce costs. 

Image Credit: Meta

………………………………………………………………………………………………………………………………………………………..

References:

OCP Summit 2025: The Open Future of Networking Hardware for AI

https://www.opencompute.org/blog/introducing-esun-advancing-ethernet-for-scale-up-ai-infrastructure-at-ocp

Networking at the Heart of AI — @Scale: Networking 2025 Recap

Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)

Gartner: AI spending >$2 trillion in 2026 driven by hyperscalers data center investments

AI Data Center Boom Carries Huge Default and Demand Risks

Analysis: Cisco, HPE/Juniper, and Nvidia network equipment for AI data centers

Qualcomm to acquire Alphawave Semi for $2.4 billion; says its high-speed wired tech will accelerate AI data center expansion

Cisco CEO sees great potential in AI data center connectivity, silicon, optics, and optical systems

Data Center Networking Market to grow at a CAGR of 6.22% during 2022-2027 to reach $35.6 billion by 2027

AT&T deploys nationwide 5G SA while Verizon lags and T-Mobile leads

Posted on by Alan Weissberger

In a blog published today (October 8th), Yigal Elbaz, AT&T’s senior VP and network CTO, AT&T  announced its 5G Standalone (SA) network is now deployed nationwide, marking an important milestone many years in the making.   Elbaz described the 5G SA nationwide deployment as “another bold leap” in wireless connectivity and said the operator is moving customers onto the network “in select areas every day.”

In fact, that “bold leap” was expected to be realized years ago! In 2021, AT&T partnered with Microsoft to offload its mobile 5G Standalone (SA) core network and other network cloud operations to Microsoft Azure, acquiring AT&T’s Network Cloud technology, software, and network operations team in the process. The goal was for Microsoft to manage the software development and deployment of AT&T’s network functions on Azure, allowing AT&T to accelerate innovation, improve efficiency, and reduce operating costs. This move has provided a strategic win for Microsoft’s Azure for Operators division by integrating AT&T’s technology and offering it to other telecom companies.

AT&T said they have millions of customers already on their 5G SA network, and we’re expanding availability to more customers as device support and provisioning allow.  Elbaz elaborated:

5G Standalone networks have now reached a level of maturity that enables our nationwide expansion. This growth is powered by an open and virtualized network, which enables us to scale efficiently and foster collaboration within an open ecosystem of partners. By embracing this open and virtualized network architecture, we are not only modernizing our infrastructure but also unlocking significant advantages for our customers and partners. This approach not only accelerates our ability to roll out new technologies like 5G Standalone but also helps ensure our customers benefit from a network that is robust, innovative, and designed with their needs in mind.

With 5G Standalone now nationwide, we’ve set the stage for the next wave of innovation, creativity, and connection. I couldn’t be prouder of our teams who made this possible, and we’ll continue to scale 5G Standalone over time and set the stage for next generation applications and services.

Compatible 5G SA smartphones include models released in the last several years starting with Apple’s iPhone 13, Samsung’s Galaxy S21 and Google’s Pixel 8.

AT&T also said its 5G Reduced Capability (RedCap) network, which uses the 5G SA core and supports the new Apple Watch Series 11, Apple Watch Ultra 3, and Apple Watch SE 3, has been expanded to 250 million points of presence.  AT&T 5G RedCap customers can look forward to a growing portfolio of devices, Elbaz said. 

……………………………………………………………………………………………………………………………………………………………………………….

Separately, Verizon is closing in on completing its 5G SA upgrade. The operator says its 5G SA is deployed nationwide but there are some places where it is still in the process of being rolled out. Although the deployment is not 100% complete, “the vast majority” of 5G SA capable phones will connect to Verizon’s network in “the vast majority of places,” according to an operator spokesperson.

Verizon has launched two network slicing services based on the 5G SA network. In April, the operator launched Frontline, a network slice for first responders that is available across the country. It also offers Enhanced Video Calling, which provides a network slice for better video communications on iPhones.

T-MobileUS launched 5G SA in 2020 and has since rolled out 5G Advanced nationwide. It also offers two network slicing propositions, T-Priority for first responders and SuperMobile for enterprise customers. Both AT&T and Verizon have implemented cloud-native 5G core networks, but T-Mobile’s implementation is more traditional. At Mobile World Congress earlier this year, T-Mobile announced its telco cloud strategy for core and edge networks that is based on Red Hat (owned by IBM).

A recent Heavy Reading (now part of Omdia) survey found 5G SA is poised to scale rapidly. Gabriel Brown noted that the results show “a critical mass is building behind 5G SA that will unlock innovation in the wider mobile network services ecosystem.”   

“This matters when it comes to layering in new services because a cloud native deployment allows operators to be more agile and deploy services faster,” Brown added.

References:

https://about.att.com/blogs/2025/5g-standalone-nationwide.html

https://about.att.com/blogs/2025/5g-redcap.html

https://www.lightreading.com/5g/at-t-verizon-chase-t-mobile-with-nationwide-5g-sa

AT&T 5G SA Core Network to run on Microsoft Azure cloud platform

Téral Research: 5G SA core network deployments accelerate after a very slow start

Building and Operating a Cloud Native 5G SA Core Network

Ookla: Europe severely lagging in 5G SA deployments and performance

Vision of 5G SA core on public cloud fails; replaced by private or hybrid cloud?

GSA: More 5G SA devices, but commercial 5G SA deployments lag

GSA 5G SA Core Network Update Report

Latest Ericsson Mobility Report talks up 5G SA networks and FWA

Global 5G Market Snapshot; Dell’Oro and GSA Updates on 5G SA networks and devices

Dell’Oro: Mobile Core Network market has lowest growth rate since 4Q 2017

5G SA networks (real 5G) remain conspicuous by their absence

 

Amazon’s Jeff Bezos at Italian Tech Week: “AI is a kind of industrial bubble”

Posted on by Alan Weissberger

Tech firms are spending hundreds of billions of dollars on advanced AI chips and data centers, not just to keep pace with a surge in the use of chatbots such as ChatGPT, Gemini and Claude, but to make sure they’re ready to handle a more fundamental and disruptive shift of economic activity from humans to machines. The final bill may run into the trillions. The financing is coming from venture capital, debt and, lately, some more unconventional arrangements that have raised concerns among top industry executives and financial asset management firms.

At Italian Tech Week in Turin on October 3, 2025, Amazon founder Jeff Bezos said this about artificial intelligence,  “This is a kind of industrial bubble, as opposed to financial bubbles.”  Bezos differentiated this from “bad” financial or housing bubbles, which  cause harm. Bezos’s comparison of the current AI boom to a historical “industrial bubble” highlights that, while speculative, it is rooted in real, transformative technology. 

“It can even be good, because when the dust settles and you see who are the winners, societies benefit from those investors,” Bezos said. “That is what is going to happen here too. This is real, the benefits to society from AI are going to be gigantic.”

He noted that during bubbles, everything (both good and bad investments) gets funded. When these periods of excitement come along, investors have a hard time distinguishing the good ideas from the bad, he said, adding this is “probably happening today” with AI investments.  “Investors have a hard time in the middle of this excitement, distinguishing between the good ideas and the bad ideas,” Bezos said of the AI industry. “And that’s also probably happening today,” he added.

  • A “good” kind of bubble: He explained that during industrial bubbles, excessive funding flows to both good and bad ideas, making it hard for investors to distinguish between them. However, the influx of capital spurs significant innovation and infrastructure development that ultimately benefits society once the bubble bursts and the strongest companies survive.
  • Echoes of the dot-com era: Bezos drew a parallel to the dot-com boom of the 1990s, where many internet companies failed, but the underlying infrastructure—like fiber-optic cable—endured and led to the creation of companies like Amazon.
  • Gigantic benefits: Despite the market frothiness, Bezos reiterated that AI is “real” and its benefits to society “are going to be gigantic.”
Bezos is not the only high-profile figure to express caution about the AI boom:
  • Sam Altman (OpenAI): The CEO of OpenAI has stated that he believes “investors as a whole are overexcited about AI.” In In August, the OpenAI CEO told reporters the AI market was in a bubble. When bubbles happen, “smart people get overexcited about a kernel of truth,” Altman warned, drawing parallels with the dot-com boom. Still, he said his personal belief is “on the whole, this would be a huge net win for the economy.”
  • David Solomon (Goldman Sachs): Also speaking at Italian Tech Week, the Goldman Sachs CEO warned that a lot of capital deployed in AI would not deliver returns and that a market “drawdown” could occur.
  • Mark Zuckerberg (Meta): Zuckerberg has also acknowledged that an AI bubble exists. The Meta CEO acknowledged that the rapid development of and surging investments in AI stands to form a bubble, potentially outpacing practical productivity and returns and risking a market crash.  However, he would rather “misspend a couple hundred billion dollars” on AI development than be late to the technology.
  • Morgan Stanley Wealth Management’s chief investment officer, Lisa Shalett, warned that the AI stock boom was showing “cracks” and was likely closer to its end than its beginning. The firm cited concerns over negative free cash flow growth among major AI players and increasing speculative investment. Shalett highlighted that free cash flow growth for the major cloud providers, or “hyperscalers,” has turned negative. This is viewed as a key signal of the AI capital expenditure cycle’s maturity. Some analysts estimate this growth could shrink by about 16% over the next year.
Image Credit:  Dreamstime.com  © Skypixel
………………………………………………………………………………………………………………………………………………………………………………………
Bezos’s remarks come as some analysts express growing fears of an impending AI market crash.
  • Underlying technology is real: Unlike purely speculative bubbles, the AI boom is driven by a fundamental technology shift with real-world applications that will survive any market correction.
  • Historical context: Some analysts believe the current AI bubble is on a much larger scale than the dot-com bubble due to the massive influx of investment.
  • Significant spending: The level of business spending on AI is already at historic levels and is fueling economic growth, which could cause a broader economic slowdown if it were to crash.
  • Potential for disruption: The AI industry faces risks such as diminishing returns for costly advanced models, increased competition, and infrastructure limitations related to power consumption. 

Ian Harnett argues, the current bubble may be approaching its “endgame.” He wrote in the Financial Times:

“The dramatic rise in AI capital expenditure by so-called hyperscalers of the technology and the stock concentration in US equities are classic peak bubble signals. But history shows that a bust triggered by this over-investment may hold the key to the positive long-run potential of AI.

Until recently, the missing ingredient was the rapid build-out of physical capital. This is now firmly in place, echoing the capex boom seen in the late-1990s bubble in telecommunications, media and technology stocks. That scaling of the internet and mobile telephony was central to sustaining ‘blue sky’ earnings expectations and extreme valuations, but it also led to the TMT bust.”

Today’s AI capital expenditure (capex) is increasingly being funded by debt, marking a notable shift from previous reliance on cash reserves. While tech giants initially used their substantial cash flows for AI infrastructure, their massive and escalating spending has led them to increasingly rely on external financing to cover costs.

This is especially true of Oracle, which will have to increase its capex by almost $100 billion over the next two years for their deal to build out AI data centers for OpenAI.  That’s an annualized growth rate of some 47%, even though Oracle’s free cash flow has already fallen into negative territory for the first time since 1990.  According to a recent note from KeyBanc Capital Markets, Oracle may need to borrow $25 billion annually over the next four years.  This comes at a time when Oracle is already carrying substantial debt and is highly leveraged. As of the end of August, the company had around $82 billion in long-term debt, with a debt-to-equity ratio of roughly 450%. By comparison, Alphabet—the parent company of Google—reported a ratio of 11.5%, while Microsoft’s stood at about 33%.  In July, Moody’s revised Oracle’s credit outlook to negative from, while affirming its Baa2 senior unsecured rating. This negative outlook reflects the risks associated with Oracle’s significant expansion into AI infrastructure, which is expected to lead to elevated leverage and negative free cash flow due to high capital expenditures. Caveat Emptor!

References:

https://fortune.com/2025/10/04/jeff-bezos-amazon-openai-sam-altman-ai-bubble-tech-stocks-investing/

https://www.ft.com/content/c7b9453e-f528-4fc3-9bbd-3dbd369041be

Can the debt fueling the new wave of AI infrastructure buildouts ever be repaid?

AI Data Center Boom Carries Huge Default and Demand Risks

Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)

Gartner: AI spending >$2 trillion in 2026 driven by hyperscalers data center investments

Will the wave of AI generated user-to/from-network traffic increase spectacularly as Cisco and Nokia predict?

Analysis: Cisco, HPE/Juniper, and Nvidia network equipment for AI data centers

RtBrick survey: Telco leaders warn AI and streaming traffic to “crack networks” by 2030

https://fortune.com/2025/09/19/zuckerberg-ai-bubble-definitely-possibility-sam-altman-collapse/

https://finance.yahoo.com/news/why-fears-trillion-dollar-ai-130008034.html

Huawei to Double Output of Ascend AI chips in 2026; OpenAI orders HBM chips from SK Hynix & Samsung for Stargate UAE project

Posted on by Alan Weissberger

With sales of Nvidia AI chips restricted in China, Huawei Technologies Inc. plans to make about 600,000 of its 910C Ascend chips next year, roughly double this year’s output, people familiar with the matter told Bloomberg. The China tech behemoth will increase its Ascend product line in 2026 to as many as 1.6 million dies – the basic silicon component that’s packaged as a chip.

Huawei had struggled to get those products to potential customers for much of 2025, because of U.S. sanctions.  Yet if Huawei and its partner Semiconductor Manufacturing International Corp. (SMIC) can hit that ambitious AI chip manufacturing target, it suggest self sufficiency which will remove some of the bottlenecks that’ve hindered not just its AI business.

The projections for 2025 and 2026 include dies that Huawei has in inventory, as well as internal estimates of yields or the rate of failure during production, the people said. Shares in SMIC and rival chipmaker Hua Hong Semiconductor Ltd. gained more than 4% in Hong Kong Tuesday, while the broader market stayed largely unchanged.

Huawei Ascend branding at a trade show i China. Photographer: Ying Tang/Getty Images

Chinese AI companies from Alibaba Group Holding Ltd. to DeepSeek need millions of AI chips to develop and operate AI services. Nvidia alone was estimated to have sold a million H20 chips in 2024.

What Bloomberg Economics Says:

Huawei’s reported plan to double AI-chip output over the next year suggests China is making real progress in working around US export controls. Yet the plan also exposes the limitations imposed by US controls: Node development remains stuck at 7 nanometers, and Huawei will continue to rely on stockpiles of foreign high-bandwidth memory amid a lack of domestic production.

From Beijing’s perspective, Huawei’s production expansion represents another move in an ongoing back-and-forth with the West over semiconductor access and self-sufficiency. The priority remains accelerating indigenization of critical technologies while steadily pushing back against Western controls.

– Michael Deng, analyst

While Huawei’s new AI silicon promises massive performance gains it has several shortcomings, especially the lack of a developer community comparable to Nvidia’s CUDA ecosystem.  A Chinese tech executive said Nvidia’s biggest advantage wasn’t its advanced chips but the ecosystem built around CUDA, its parallel computing architecture and programming model. The exec called for the creation of a Chinese version of CUDA that can be used worldwide. 

Also, Huawei is playing catchup by progressively going open source. It announced last month that its Ascend and AI training toolkit CANN, its Mind development environment and Pangu models would all be open source by year-end.

Huawei chairman Eric Xu said in an interview the company had given the “ecosystem issue” a great deal of thought and regarded the transition to open source as a long-term project. “Why keep it hidden? If it’s widely used, an ecosystem will emerge; if it’s used less, the ecosystem will disappear,” he said.

………………………………………………………………………………………………………………………………………………………………………

At its customer event in Shanghai last month, Huawei revealed that it planned to spend 15 billion Chinese yuan (US$2.1 billion) annually over the next five years on ecosystem development and open source computing.

Xu announced a series of new Ascend chips – the 950, 960 and 970 – to be rolled out over the next three years.  He foreshadowed a new series of massive Atlas SuperPoD clusters – each one a single logical machine made up of multiple physical devices that can work together – and also announced Huawei’s unified bus interconnect protocol, which allows customers to stitch together compute power across multiple data centers. 

Xu acknowledged that Huawei’s single Ascend chips could not match Nvidia’s, but said the SuperPoDs were currently the world’s most powerful and will remain so “for years to come.” But the scale of its SuperPOD architecture points to its other shortcoming – the power consumption of these giant compute arrays. 

………………………………………………………………………………………………………………………………………………………………………….

Separately, OpenAI has made huge memory chip agreements with South Korea’s SK Hynix and Samsung, the world’s two biggest semiconductor memory manufacturers.  The partnership, aimed at locking up HBM ((High Bandwidth Memory) [1.] chip supply for the $400 billion Stargate AI infrastructure project, is estimated to be worth more than 100 trillion Korean won (US$71.3 billion) for the Korean chipmakers over the next four years. The two companies say they are targeting 900,000 DRAM wafer starts per month – more than double the current global HBM capacity.

Note 1. HBM is a specialized type of DRAM that uses a unique 3D vertical stacking architecture and Through-Silicon Via (TSV) technology to achieve significantly higher bandwidth and performance than traditional, flat DRAM configurations. HBM uses standard DRAM “dies” stacked vertically, connected by TSVs, to create a more densely packed, high-performance memory solution for demanding applications like AI and high-performance computing.

…………………………………………………………………………………………………………………………………………………………………………….

“These partnerships will focus on increasing the supply of advanced memory chips essential for next-generation AI and expanding data center capacity in Korea, positioning Samsung and SK as key contributors to global AI infrastructure and supporting Korea’s ambition to become a top-three global AI nation.” OpenAI said.

The announcement followed a meeting between President Lee Jae-myung, Samsung Electronics Executive Chairman Jay Y. Lee, SK Chairman Chey Tae-won, and OpenAI CEO Sam Altman at the Presidential Office in Seoul.

Through these partnerships, Samsung Electronics and SK hynix plan to scale up production of advanced memory chips, targeting 900,000 DRAM wafer starts per month at an accelerated capacity rollout, critical for powering OpenAI’s advanced AI models.

OpenAI also signed a series of agreements today to explore developing next-generation AI data centers in Korea. These include a Memorandum of Understanding (MoU) with the Korean Ministry of Science and ICT (MSIT) specifically to evaluate opportunities for building AI data centers outside the Seoul Metropolitan Area, supporting balanced regional economic growth and job creation across the country.

The agreements signed today also include a separate partnership with SK Telecom to explore building an AI data center in Korea, as well as an agreement with Samsung C&T, Samsung Heavy Industries, and Samsung SDS to assess opportunities for additional data center capacity in the country.

References:

https://www.bloomberg.com/news/articles/2025-09-29/huawei-to-double-output-of-top-ai-chip-as-nvidia-wavers-in-china

https://www.lightreading.com/ai-machine-learning/huawei-sets-itself-as-china-s-go-to-for-ai-tech

https://openai.com/index/samsung-and-sk-join-stargate/

OpenAI orders $71B in Korean memory chips

AI Data Center Boom Carries Huge Default and Demand Risks

Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)

U.S. export controls on Nvidia H20 AI chips enables Huawei’s 910C GPU to be favored by AI tech giants in China

Huawei launches CloudMatrix 384 AI System to rival Nvidia’s most advanced AI system

China gaining on U.S. in AI technology arms race- silicon, models and research

Will billions of dollars big tech is spending on Gen AI data centers produce a decent ROI?

Can the debt fueling the new wave of AI infrastructure buildouts ever be repaid?

Despite U.S. sanctions, Huawei has come “roaring back,” due to massive China government support and policies

Can the debt fueling the new wave of AI infrastructure buildouts ever be repaid?

Posted on by Alan Weissberger

IEEE Techblog has called attention to the many challenges and risks inherent in the current mega-spending boom for AI infrastructure (building data centers, obtaining power/electricity, cooling, maintenance, fiber optic networking, etc) .  In particular, these two recent blog posts:

AI Data Center Boom Carries Huge Default and Demand Risks and

Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)

This article focuses on the tremendous debt that Open AI, Oracle and newer AI cloud companies will have to obtain and the huge hurdles they face to pay back the money being spent to build out their AI infrastructures. While the major hyperscalers (Amazon, Microsoft, Google and Meta) are in good financial shape and won’t need to take on much debt, a  new wave of  heavily leveraged firms is emerging—one that could reshape the current AI boom.

OpenAI, for example, is set to take borrowing and large-scale contracts to an unbelievable new level. OpenAI is planning a vast network of data centers expected to cost at least $1 trillion over the coming years. As part of this effort, the company signed a $300 billion, five-year contract this month under which Oracle “is to set up AI computing infrastructure and lease it to OpenAI.”   In other words, OpenAI agreed to pay Oracle $300 billion over five years for the latter company to build out new AI data centers.  Where will OpenAI get that money?  It will be be burning billions in cash and won’t be profitable till 2029 at the earliest.

To fulfill its side of the deal, Oracle will need to invest heavily in infrastructure before receiving full payment—requiring significant borrowing. According to a recent note from KeyBanc Capital Markets, Oracle may need to borrow $25 billion annually over the next four years.  This comes at a time when Oracle is already carrying substantial debt and is highly leveraged. As of the end of August, the company had around $82 billion in long-term debt, with a debt-to-equity ratio of roughly 450%. By comparison, Alphabet—the parent company of Google—reported a ratio of 11.5%, while Microsoft’s stood at about 33%.

Companies like Oracle and other less-capitalized AI players such as CoreWeave have little choice but to take on more debt if they want to compete at the highest level. Nebius Group, another Nasdaq-listed AI cloud provider similar to CoreWeave, struck a $19.4 billion deal in September to provide AI computing services to Microsoft. The company announced it would finance the necessary capital expenditures “through a combination of its cash flow and debt secured against the contract.”

………………………………………………………………………………………………………………………………………………………………………………………………

Sidebar – Stock market investors seem to love debt and risk:

CoreWeave’s shares have more than tripled since its IPO in March, while Nebius stock jumped nearly 50% after announcing its deal with Microsoft. Not to be outdone, Oracle’s stock surged 40% in a single day after the company disclosed a major boost in projected revenue from OpenAI in its infrastructure deal—even though the initiative will require years of heavy spending by Oracle.

–>What’s so amazing to this author is that OpenAI selected Oracle for the AI infrastructure it will use, even though the latter is NOT a major cloud service provider and is certainly not a hyperscaler.  For Q1 2025, it held about 3% market share, placing it #5 among global cloud service providers.

…………………………………………………………………………………………………………………………………………………………………………………………………

Data Center Compute Server & Storage Room;  iStock Photo credit: Andrey Semenov

……………………………………………………………………………………………………………………………………………….

Among other new AI Cloud players:

  • CyrusOne secured nearly $12 billion in financing (much in debt) for AI / data center expansion. Around $7.9 billion of that is for new data center / AI digital infrastructure projects in the U.S.
  • SoftBank / “Stargate” initiative: The Stargate project (OpenAI + Oracle + SoftBank + MGX, etc.) is being structured with major debt. The plan is huge—around $500 billion in AI infrastructure and supercomputers, and financing is expected to be ~70% debt, ~10% equity among the sources.
  • xAI (Elon Musk’s AI firm):  xAI raised $10 billion in combined debt + equity. Specifically ~$5 billion in secured notes / term loans (debt), with the remainder in equity. The money is intended to build out its AI infrastructure (e.g. GPU facilities / data centers).

There’s growing skepticism about whether these companies can meet their massive contract obligations and repay their debts. Multiple recent studies suggest AI adoption isn’t advancing as quickly as supporters claim. One study found that only 3% of consumers are paying for AI services. Forecasts projecting trillions of dollars in annual spending on AI data centers within a few years appear overly optimistic.

OpenAI’s position, despite the hype, seems very shaky. D.A. Davidson analyst Gil Luria estimates the company would need to generate over $300 billion in annual revenue by 2030 to justify the spending implied in its Oracle deal—a steep climb from its current run rate of about $12 billion. OpenAI has financial backing from SoftBank and Nvidia, with Nvidia pledging up to $100 billion, but even that may not be enough.  “A vast majority of Oracle’s data center capacity is now promised to one customer, OpenAI, who itself does not have the capital to afford its many obligations,” Luria said.

Oracle could try to limit risk by pacing its spending with revenue received from OpenAI.  Nonetheless, Moody’s flagged “significant” risks in a recent note, citing the huge costs of equipment, land, and electricity. “Whether these will be financed through traditional debt, leases or highly engineered financing vehicles, the overall growth in balance sheet obligations will also be extremely large,” Moody’s warned. In July (two months before the OpenAI deal), it gave Oracle a negative credit outlook.

There’s a real possibility that things go smoothly. Oracle may handle its contracts and debt well, as it has in the past. CoreWeave, Nebius, and others might even pioneer new financial models that help accelerate AI development.

It’s very likely that some of today’s massive AI infrastructure deals will be delayed, renegotiated, or reassigned if AI demand doesn’t grow as fast as AI spending. Legal experts say contracts could be transferred.  For example, if OpenAI can’t make the promised, Oracle might lease the infrastructure to a more financially stable company, assuming the terms allow it.

Such a shift wouldn’t necessarily doom Oracle or its debt-heavy peers. But it would be a major test for an emerging financial model for AI—one that’s starting to look increasingly speculative.  Yes, even bubbly!

………………………………………………………………………………………………………………………………………………………………………………

References:

https://www.wsj.com/tech/ai/debt-is-fueling-the-next-wave-of-the-ai-boom-278d0e04

https://www.crn.com/news/cloud/2025/cloud-market-share-q1-2025-aws-dips-microsoft-and-google-show-growth

AI Data Center Boom Carries Huge Default and Demand Risks

Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)

Should Peak Data Rates be specified for 5G (IMT 2020) and 6G (IMT 2030) networks?

Posted on by Alan Weissberger

Peak Data Rate [1.] is one of the most visible attributes of IMT (International Mobile Telecommunications) cellular networks, e.g. 3G, 4G and 5G. As a result, it gets significant attention from analysts and reporters that create high expectations for  IMT end users which may never be realized in commercially deployed IMT networks.

For example, the peak data rates specified by the ITU-R M.2410 report for IMT-2020 (5G) have not been realized in any 5G production networks under typical conditions. The ITU-R’s 20 Gbps downlink and 10 Gbps uplink targets are theoretical maximums, achievable only in a controlled test environment with ideal conditions. Please refer to the chart below.

……………………………………………………………………………………………………………………………………………………………………..

Note 1. Peak data rate is the theoretical maximum [achievable] data rate under ideal conditions, which is the received data bits assuming error-free conditions assignable to a single mobile station, when all assignable radio resources for the corresponding link direction are utilized (i.e. excluding radio resources that are used for physical layer synchronization, reference signals or pilots, guard bands and guard times).

………………………………………………………………………………………………………………………………………………………………………

5G services are deployed across three main frequency ranges and the speed capability varies dramatically for each.

  • Low-band (sub-6 GHz): Offers wide coverage but only a modest speed improvement over 4G, typically delivering a few hundred Mbps at best.
  • Mid-band (sub-6 GHz): Provides a balance of speed and coverage, with peak speeds sometimes reaching 1 Gbps, though typical average speeds are much lower.
  • High-band (millimeter wave or mmWave): This is the only band capable of reaching multi-gigabit speeds. However, its signal range is very short and it is easily blocked by physical objects, limiting its availability to dense urban areas and specific venues.  5G mmWave base station power consumption is also very high which limits coverage.
Several factors are critical for pushing the boundaries of 5G downlink speeds in live networks:
  • mmWave spectrum: Higher-band millimeter wave spectrum offers massive bandwidth, enabling multi-gigabit speeds. However, its use is limited to dense urban areas and specific venues due to its short range.
  • Carrier aggregation: Combining multiple frequency bands (e.g., mmWave with mid-band) significantly increases the total available bandwidth and is crucial for achieving the highest download speeds.
  • 5G Advanced (Release 18): New developments in 5G-Advanced technology (also known as 5.5G) enable even higher performance. The Telstra record in 2025 utilized 5G Advanced software.
  • Equipment and device capabilities: Peak speeds require cutting-edge network hardware from vendors like Ericsson, Nokia, and Samsung, as well as the latest mobile devices powered by advanced modems from companies like Qualcomm and MediaTek.

The gap between what IMT-2020 (5G) technology can deliver (on paper) and what is actually realized in commercial 5G networks  has grown larger and larger over the past few years [2.].  Here’s a summary of speed differences:

Speed metric ITU-R specification Reality in commercial networks
Peak data rate 20 Gbps (downlink)

10 Gbps (uplink)

 Reached only in isolated demonstrations, typically using high-band mmWave technology.
User experienced rate 100 Mbps (downlink)

15 to 50 Mbps (uplink)

 The typical average speed for many users, particularly on low- and mid-band deployments.  mmWave is higher, but the range is limited.

Note 2.  The gap is even greater for 5G latency! The minimum required latency in ITU-R M.2410 for user plane are:
– 4 ms for eMBB
– 1 ms for URLLC
assumes unloaded conditions (a single user) for small IP packets (e.g. 0 byte payload + IP header), for both downlink and uplink.

The minimum requirement for control plane latency is 20 ms. Proponents are encouraged to consider lower control plane latency, e.g. 10 ms.

However, the average latency experienced in deployed commercial 5G networks is higher, typically ranging between 5 and 20 milliseconds, depending on the network architecture, spectrum, and use case.  One reason is that the 3GPP Release 16 spec for 5G-NR enhancements for URLLC in the RAN and Core network were never completed.

5G mmWave spectrum has the potential for the lowest latency, but its limited range and line-of-sight requirements limit restrict deployments to dense urban areas.  Therefore, most 5G users connect via mid-band or low-band, which have higher latency.

……………………………………………………………………………………………………………………………………………………………….

For that reason, several companies (Apple, Nokia, TELECOM ITALIA, Deutsche Telekom, SK Telecom, Spark NZ, AT&T) have proposed not to define IMT-2030 peak data rate requirement values in ITU-R M.[IMT-2030.TECH PERF REQ] nor to maintain the IMT-2020 (5G) peak data rate numbers from the ITU-R M.2410 report.

Author’s Note: The IMT-2030 performance requirements in ITU-R M.[IMT-2030.TECH PERF REQ] are to be evaluated according to the criteria defined in Report ITU-R M.[IMT‑2030.EVAL] and Report ITU-R M.[IMT-2030.SUBMISSION] for the development of IMT-2030 recommendations within ITU-R WP5D.

……………………………………………………………………………………………………………………………………………………………………………….

Addendum – Measurements of top 5G network speeds:

  • In the first half of 2025, Ookla said  e& in the United Arab Emirates was the world’s fastest 5G network, noting a median upload speed of 52.21 Mbps. Other top performers like South Korea, Qatar, and Brazil also see median speeds well above 20 Mbps.
  • U.S. performance: In the U.S., major carriers are in a close race. In mid-2024, Opensignal found Verizon with the fastest 5G upload speed at 21.2 Mbps, with T-Mobile close behind. However, as of early 2025, a separate Opensignal report credited T-Mobile with the fastest overall upload experience, at 17.9 Mbps, though that figure includes both 4G and 5G connections.
  • European performance: Speeds vary across Europe. Ookla reported that in the first half of 2025, Magenta Telekom in Austria achieved a median 5G upload speed of 35.67 Mbps, while Three in the U.K. recorded a median of 13.07 Mbps.
  • Rural vs. urban divide: Average 5G uplink speeds are often higher in urban areas where mid-band spectrum is more prevalent. However, as of mid-2023, Opensignal noted that the rural-urban gap for 5G upload speeds in the U.S. was narrowing due to increased rural investment.
  • Dependence on network type: Whether a network uses 5G standalone (SA) or non-standalone (NSA) architecture impacts speeds. In early 2025, an analysis in the U.K. showed that while 5G SA had lower latency, 5G NSA still had a slightly higher proportion of high-speed uplink connections. 

…………………………………………………………………………………………………………………………………………………………

References:

https://www.itu.int/en/ITU-R/study-groups/rsg5/rwp5d/imt-2020/Documents/S01-1_Requirements%20for%20IMT-2020_Rev.pdf

https://www.itu.int/pub/r-rep-m.2410-2017

https://www.itu.int/dms_pub/itu-r/opb/rep/R-REP-M.2410-2017-PDF-E.pdfITU-R WP 5D reports on: IMT-2030 (“6G”) Minimum Technology Performance Requirements; Evaluation Criteria & Methodology

3GPP Release 16 5G NR Enhancements for URLLC in the RAN & URLLC in the 5G Core network

 

IMT-2030 Technical Performance Requirements (TPR) from ITU-R WP5D

Key Objectives of WG Technology Aspects at ITU-R WP 5D meeting June 24-July 3, 2025

ITU-R WP5D IMT 2030 Submission & Evaluation Guidelines vs 6G specs in 3GPP Release 20 & 21

ITU-R: IMT-2030 (6G) Backgrounder and Envisioned Capabilities

Draft new ITU-R recommendation (not yet approved): M.[IMT.FRAMEWORK FOR 2030 AND BEYOND]

ITU-R M.2150-1 (5G RAN standard) will include 3GPP Release 17 enhancements; future revisions by 2025

 

 

Aviat Networks and Intracom Telecom partner to deliver 5G mmWave FWA in North America

Posted on by Alan Weissberger

Aviat Networks, a wireless transport and access company, today announced a partnership with Intracom Telecom, a global technology systems and solutions provider, to deliver Fixed Wireless Access (FWA) technology using high-capacity 28 and 39 GHz millimeter wave (mmWave) bands, conforming to FCC requirements for mmWave bands intended for 5G use.

Aviat will initially focus on select North American service providers to address the growing need for multi-Gigabit consumer and enterprise 5G use cases as an alternative to the high cost, delays and complexity of using fiber, but with fiber-like performance. In addition, Aviat will offer software solutions along with a comprehensive set of design, planning, deployment and support services thanks to its extensive presence in North America.

Intracom Telecom’s WiBAS G5 platform is the only commercially available point-to-multipoint FWA solution operating in the 28 and 39 GHz mmWave bands that can address the growing demand for high-capacity Fixed Wireless Access, cost effectively delivering over 22Gbps from the same base station site, using Multi-User MIMO and Hybrid Massive Beamforming, over distances of up to 5 miles and more.

“We are very excited at this significant opportunity to extend our wireless expertise to provide advanced mmWave FWA solutions,” Pete Smith, CEO of Aviat Networks said, “Wireless can be deployed rapidly and cost effectively, and is perfectly suited to support high speed connectivity combined with excellent reliability.”

“I am very proud of Intracom Telecom’s R&D team for creating a solution that sets a new benchmark for FWA. Through this strategic partnership with Aviat Networks, we’re excited to help U.S. operators accelerate broadband expansion and deliver a true multi-gigabit experience, and more, over wireless,” said Kartlos Edilashvili, CEO of Intracom Telecom.

Image Source:  Aviat Networks

In the U.S., Verizon, AT&T, and T-Mobile (including UScellular‘s retail wireless operations) use 28 GHz and 39 GHz millimeter wave (mmWave) bands for 5G services  in densely populated areas and venues. mmWave signal propagation characteristics limit range/coverage and have a high susceptibility to blockage by physical objects and weather. These limitations significantly increase deployment costs and constrain coverage to densely populated areas. 

About Aviat Networks:

Aviat, based in Austin, TX, is a leading expert in wireless transport and access solutions and works to provide dependable products, services and support to its customers. With more than one million systems sold into 170 countries worldwide, communications service providers and private network operators including state/local government, utility, federal government and defense organizations trust Aviat with their critical applications. Coupled with a long history of microwave innovations, Aviat provides a comprehensive suite of localized professional and support services enabling customers to drastically simplify both their networks and their lives. For more than 70 years, the experts at Aviat have delivered high performance products, simplified operations, and the best overall customer experience. Aviat is headquartered in Austin, Texas. For more information, visit www.aviatnetworks.com or connect with Aviat Networks on LinkedIn and Facebook.

About Intracom Telecom:

Intracom Telecom is a global technology systems and solutions provider operating for over 45 years in the market. The company is the benchmark in fixed wireless access, and it successfully innovates in the wireless access & transmission field. Furthermore, the company offers a comprehensive software solutions portfolio and a complete range of ICT services. Intracom Telecom serves telecom operators, public authorities and large public and private enterprises. The Group maintains its own R&D and production facilities, operates subsidiaries worldwide and has been active in the North American market since 2001, through its subsidiary, Intracom Telecom USA, based in Atlanta, Georgia. The parent company is located in Athens, Greece. For more information, visit www.intracom-telecom.com

References:

https://www.prnewswire.com/news-releases/aviat-networks-and-intracom-telecom-partnership-delivers-fixed-wireless-access-fwa-solutions-302566564.html

T-Mobile’s growth trajectory increases: 5G FWA, Metronet acquisition and MVNO deals with Charter & Comcast

Dell’Oro: 4G and 5G FWA revenue grew 7% in 2024; MRFR: FWA worth $182.27B by 2032

Latest Ericsson Mobility Report talks up 5G SA networks and FWA

Aviat Sells TIP compliant 5G-Ready Disaggregated Transmission Network to Africell

Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)

Posted on by Alan Weissberger

Big Tech plans to spend between $364 billion and $400 billion on AI data centers, purchasing specialized AI hardware like GPUs, and supporting cloud computing/storage capacity. The final 2Q 2025 GDP report, released last week, reveals a surge in data center infrastructure spending from $9.5 billion in early 2020 to $40.4 billion in the second quarter of 2025.  It’s largely due to an unprecedented investment boom driven by artificial intelligence (AI) and cloud computing. The increase highlights a monumental shift in capital expenditure by major tech companies.

Yet there are huge uncertainties about how far AI will transform scientific discovery and hypercharge technological advance.  Tech financial analysts worry that enthusiasm for AI has turned into a bubble that is reminiscent of the mania around the internet’s infrastructure build-out boom from 1998-2000.  During that time period, telecom network providers spent over $100 billion blanketing the country with fiber optic cables based on the belief that the internet’s growth would be so explosive that such massive investments were justified.  The “talk of the town” during those years was the “All Optical Network,” with ultra-long haul optical transceiver, photonic switches and optical add/drop multiplexers.  27 years later, it still has not been realized anywhere in the world.

The resulting massive optical network overbuilding  made telecom the hardest hit sector of the dot-com bust. Industry giants toppled like dominoes, including Global Crossing, WorldCom, Enron, Qwest, PSI Net and 360Networks.

However, a key difference between then and now is that today’s tech giants (e.g. hyperscalers) produce far more cash than the fiber builders in the 1990s. Also, AI is immediately available for use by anyone that has a high speed internet connection (via desktop, laptop, tablet or smartphone) unlike the late 1990s when internet users (consumers and businesses) had to obtain high-speed wireline access via cable modems, DSL or (in very few areas) fiber to the premises.

……………………………………………………………………………………………………………………………………………………………………………………………………………………………….

Today, the once boring world of chips and data centers has become a raging multi-hundred billion dollar battleground where Silicon Valley giants attempt to one up each other with spending commitments—and sci-fi names.   Meta CEO Mark Zuckerberg teased his planned “Hyperion” mega-data center with a social-media post showing it would be the size of a large chunk of Manhattan.

OpenAI’s Sam Altman calls his data-center effort “Stargate,” a reference to the 1994 movie about an interstellar time-travel portal. Company executives this week laid out plans that would require at least $1 trillion in data-center investments, and Altman recently committed the company to pay Oracle an average of approximately $60 billion a year for AI compute servers in data centers in coming years. That’s despite Oracle is not a major cloud service provider and OpenAI will not have the cash on hand to pay Oracle.

In fact, OpenAI is on track to realize just $13 billion in revenue from all its paying customers this year and won’t be profitable till at least 2029 or 2030. The company projects its total cash burn will reach $115 billion by 2029.  The majority of its revenue comes from subscriptions to premium versions of ChatGPT, with the remainder from selling access to its models via its API. Although ~ 700 million people—9% of the world’s population—are weekly users of ChatGPT (as of August, up from 500 million in March), its estimated that over 90% use the free version.  Also this past week:

  • Nvidia plans to invest up to $100 billion to help OpenAI build data center capacity with millions GPUs.
  • OpenAI revealed an expanded deal with Oracle and SoftBank , scaling its “Stargate” project to a $400 billion commitment across multiple phases and sites.
  • OpenAI deepened its enterprise reach with a formal integration into Databricks — signaling a new phase in its push for commercial adoption.

Nvidia is supplying capital and chips. Oracle is building the sites. OpenAI is anchoring the demand. It’s a circular economy that could come under pressure if any one player falters. And while the headlines came fast this week, the physical buildout will take years to deliver — with much of it dependent on energy and grid upgrades that remain uncertain.

Another AI darling is CoreWeave, a company that provides GPU-accelerated cloud computing platforms and infrastructure.  From its founding in 2017 until its pivot to cloud computing in 2019, Corweave was an obscure cryptocurrency miner with fewer than two dozen employees. Flooded with money from Wall Street and private-equity investors, it has morphed into a computing goliath with a market value bigger than General Motors or Target.

…………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………….

Massive AI infrastructure spending will require tremendous AI revenue for pay-back:

David Cahn, a partner at venture-capital firm Sequoia, estimates that the money invested in AI infrastructure in 2023 and 2024 alone requires consumers and companies to buy roughly $800 billion in AI products over the life of these chips and data centers to produce a good investment return. Analysts believe most AI processors have a useful life of between three and five years.

This week, consultants at Bain & Co. estimated the wave of AI infrastructure spending will require $2 trillion in annual AI revenue by 2030. By comparison, that is more than the combined 2024 revenue of Amazon, Apple, Alphabet, Microsoft, Meta and Nvidia, and more than five times the size of the entire global subscription software market.

Morgan Stanley estimates that last year there was around $45 billion of revenue for AI products. The sector makes money from a combination of subscription fees for chatbots such as ChatGPT and money paid to use these companies’ data centers.  How the tech sector will cover the gap is “the trillion dollar question,” said Mark Moerdler, an analyst at Bernstein. Consumers have been quick to use AI, but most are using free versions, Moerdler said. Businesses have been slow to spend much on AI services, except for the roughly $30 a month per user for Microsoft’s Copilot or similar products. “Someone’s got to make money off this,” he said.

…………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………….

Why this time is different (?):

AI cheerleaders insist that this boom is different from the dot-com era. If AI continues to advance to the point where it can replace a large swath of white collar jobs, the savings will be more than enough to pay back the investment, backers argue. AI executives predict the technology could add 10% to global GDP in coming years.

“Training AI models is a gigantic multitrillion dollar market,” Oracle chairman Larry Ellison told investors this month. The market for companies and consumers using AI daily “will be much, much larger.”

The financing behind the AI build-out is complex. Debt is layered on at nearly every level.  his “debt-fueled arms race” involves large technology companies, startups, and private credit firms seeking innovative ways to fund the development of data centers and acquire powerful hardware, such as Nvidia GPUs. Debt is layered across different levels of the AI ecosystem, from the large tech giants to smaller cloud providers and specialized hardware firms. 

Alphabet, Microsoft, Amazon, Meta and others create their own AI products, and sometimes sell access to cloud-computing services to companies such as OpenAI that design AI models. The four “hyperscalers” alone are expected to spend nearly $400 billion on capital investments next year, more than the cost of the Apollo space program in today’s dollars.  Some build their own data centers, and some rely on third parties to erect the mega-size warehouses tricked out with cooling equipment and power.

…………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………….

Echoes of bubbles past:

History is replete with technology bubbles that pop. Optimism over an invention—canals, electricity, railroads—prompts an investor stampede premised on explosive growth. Overbuilding follows, and investors eat giant losses, even when a new technology permeates the economy.  Predicting when a boom turns into a bubble is notoriously hard. Many inflate for years. Some never pop, and simply stagnate.

The U.K.’s 19th-century railway mania was so large that over 7% of the country’s GDP went toward blanketing the country with rail. Between 1840 and 1852, the railway system nearly quintupled to 7,300 miles of track, but it only produced one-fourth of the revenue builders expected, according to Andrew Odlyzko,PhD, an emeritus University of Minnesota mathematics professor who studies bubbles. He calls the unbridled optimism in manias “collective hallucinations,” where investors, society and the press follow herd mentality and stop seeing risks.

He knows from firsthand experience as a researcher at Bell Labs in the 1990s. Then, telecom giants and upstarts raced to speculatively plunge tens of millions of miles of fiber cables into the ground, spending the equivalent of around 1% of U.S. GDP over half a decade.

Backers compared the effort to the highway system, to the advent of electricity and to discovering oil. The prevailing belief at the time, he said, was that internet use was doubling every 100 days. But in reality, for most of the 1990s boom, traffic doubled every year, Odlyzko found.

The force of the mania led executives across the industry to focus on hype more than unfavorable news and statistics, pouring money into fiber until the bubble burst.

“There was a strong element of self interest,” as companies and executives all stood to benefit financially as long as the boom continued, Odlyzko said. “Cautionary signs are disregarded.”

Kevin O’Hara, a co-founder of upstart fiber builder Level 3, said banks and stock investors were throwing money at the company, and executives believed demand would rocket upward for years. Despite worrying signs, executives focused on the promise of more traffic from uses like video streaming and games.

“It was an absolute gold rush,” he said. “We were spending about $110 million a week” building out the network.

When reality caught up, Level 3’s stock dropped 95%, while giants of the sector went bust. Much of the fiber sat unused for over a decade. Ultimately, the growth of video streaming and other uses in the early 2010s helped soak up much of the oversupply.

Worrying signs:

There are growing, worrying signs that the optimism about AI won’t pan out.

  • MIT Media Lab (2025): The “State of AI in Business 2025” report found that 95% of custom enterprise AI tools and pilots fail to produce a measurable financial impact or reach full-scale production. The primary issue identified was a “learning gap” among leaders and organizations, who struggle to properly integrate AI tools and redesign workflows to capture value.
  • A University of Chicago economics paper found AI chatbots had “no significant impact on workers’ earnings, recorded hours, or wages” at 7,000 Danish workplaces.
  • Gartner (2024–2025): The research and consulting firm has reported that 85% of AI initiatives fail to deliver on their promised value. Gartner also predicts that by the end of 2025, 30% of generative AI projects will be abandoned after the proof-of-concept phase due to issues like poor data quality, lack of clear business value, and escalating costs.
  • RAND Corporation (2024): In its analysis, RAND confirmed that the failure rate for AI projects is over 80%, which is double the failure rate of non-AI technology projects. Cited obstacles include cost overruns, data privacy concerns, and security risks.

OpenAI’s release of ChatGPT-5 in August was widely viewed as an incremental improvement, not the game-changing thinking machine many expected. Given the high cost of developing it, the release fanned concerns that generative AI models are improving at a slower pace than expected.  Each new AI model—ChatGPT-4, ChatGPT-5—costs significantly more than the last to train and release to the world, often three to five times the cost of the previous, say AI executives. That means the payback has to be even higher to justify the spending.

Another hurdle: The chips in the data centers won’t be useful forever. Unlike the dot-com boom’s fiber cables, the latest AI chips rapidly depreciate in value as technology improves, much like an older model car.  And they are extremely expensive.

“This is bigger than all the other tech bubbles put together,” said Roger McNamee, co-founder of tech investor Silver Lake Partners, who has been critical of some tech giants. “This industry can be as successful as the most successful tech products ever introduced and still not justify the current levels of investment.”

Other challenges include the growing strain on global supply chains, especially for chips, power and infrastructure. As for economy-wide gains in productivity, few of the biggest listed U.S. companies are able to describe how AI was changing their businesses for the better. Equally striking is the minimal euphoria some Big Tech companies display in their regulatory filings. Meta’s 10k form last year reads: “[T]here can be no assurance that the usage of AI will enhance our products or services or be beneficial to our business, including our efficiency or profitability.” That is very shaky basis on which to conduct a $300bn capex splurge.

………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………

Conclusions:

Big tech spending on AI infrastructure has been propping up the U.S. economy, with some projections indicating it could fuel nearly half of the 2025 GDP growth. However, this contribution primarily stems from capital expenditures, and the long-term economic impact is still being debated.  George Saravelos of Deutsche Bank notes that economic growth is not coming from AI itself but from building the data centers to generate AI capacity.

Once those AI factories have been built, with needed power supplies and cooling, will the productivity gains from AI finally be realized? How globally disseminated will those benefits be?  Finally, what will be the return on investment (ROI) for the big spending AI companies like the hyperscalers, OpenAI and other AI players?

………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………

References:

https://www.wsj.com/tech/ai/ai-bubble-building-spree-55ee6128

https://www.ft.com/content/6c181cb1-0cbb-4668-9854-5a29debb05b1

https://www.cnbc.com/2025/09/26/openai-big-week-ai-arms-race.html

https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers

Gartner: AI spending >$2 trillion in 2026 driven by hyperscalers data center investments

AI Data Center Boom Carries Huge Default and Demand Risks

AI spending is surging; companies accelerate AI adoption, but job cuts loom large

Will billions of dollars big tech is spending on Gen AI data centers produce a decent ROI?

Canalys & Gartner: AI investments drive growth in cloud infrastructure spending

AI wave stimulates big tech spending and strong profits, but for how long?

AI Echo Chamber: “Upstream AI” companies huge spending fuels profit growth for “Downstream AI” firms

OpenAI partners with G42 to build giant data center for Stargate UAE project

Big Tech and VCs invest hundreds of billions in AI while salaries of AI experts reach the stratosphere

Superclusters of Nvidia GPU/AI chips combined with end-to-end network platforms to create next generation data centers

Analysis: Cisco, HPE/Juniper, and Nvidia network equipment for AI data centers

Networking chips and modules for AI data centers: Infiniband, Ultra Ethernet, Optical Connections

OpenAI and Broadcom in $10B deal to make custom AI chips

Lumen Technologies to connect Prometheus Hyperscale’s energy efficient AI data centers

Proposed solutions to high energy consumption of Generative AI LLMs: optimized hardware, new algorithms, green data centers

Liquid Dreams: The Rise of Immersion Cooling and Underwater Data Centers

Lumen: “We’re Building the Backbone for the AI Economy” – NaaS platform to be available to more customers

Initiatives and Analysis: Nokia focuses on data centers as its top growth market

SK Telecom forms AI CIC in-house company to pursue internal AI innovation

Posted on by Alan Weissberger

SK Telecom (SKT) is establishing an in-house independent company (CIC) that consolidates its artificial intelligence (AI) capabilities. Through AI CIC, SK Telecom plans to invest approximately 5 trillion won (US$3.5 billion) in AI over the next five years and achieve annual sales of over 5 trillion won ($3.5 billion) by 2030.

On September 25th, SK Telecom CEO Ryu Young-sang held a town hall meeting for all employees at the SKT Tower Supex Hall in Jung-gu, Seoul, announcing the launch of AI CIC to pursue rapid AI innovation. Ryu will concurrently serve as the CEO of SK CIC. SK Telecom plans to unveil detailed organizational restructuring plans for AI CIC at the end of October this year.

“We are launching AI CIC, a streamlined organizational structure, and will simultaneously pursue internal AI innovation, including internal systems, organizational culture, and enhancing employees’ AI capabilities. We will grow AI CIC to be the main driver of SK’s AI business and, furthermore, the core that leads the AI business for the entire SK Group.  The AI CIC will establish itself as South Korea’s leading AI business operator in all fields of AI, including services, platforms, AI data centers and proprietary foundation models,” Ryu said.

The newly established AI CIC will be responsible for all the company’s AI-related functions and businesses. It is expected that SK Telecom’s business will be divided into mobile network operations (MNO) and AI, with AI CIC consolidating related businesses to enhance operational efficiency. Furthermore, AI CIC will actively participate in government-led AI projects, contributing to the establishment of a government-driven AI ecosystem. SKT said that reorganizing its services under one umbrella will “drive AI innovation that enhance business productivity and efficiency.”

“Through this (AI CIC), we will play a central role in building a domestic AI-related ecosystem and become a company that contributes to the success of the national AI strategy,” Ryu said.

By integrating and consolidating dispersed AI technology assets, SKT plans to strengthen the role of the “AI platform” that supports AI technology/operations across the entire SK Group, including SKT, and also pursue a strategy to secure a flexible “AI model” to respond to the diverse AI needs of the government, industry, and private sectors.

In addition, SKT will accelerate the development of future growth areas (R&D) such as digital twins and robots, and the expansion of domestic and international partnerships based on AI full-stack capabilities.

Ryu Young-sang, CEO of SK Telecom, unveils the plans for the AI CIC 

CEO Ryu said, “SK Telecom has secured various achievements such as securing 10 million Adot (AI enabled) subscribers, selecting an independent AI foundation model, launching the Ulsan AI DC, and establishing global partnerships through its transformation into an AI company over the past three years, and has laid the foundation for future leaps forward.  We will achieve another AI innovation centered around the AI ​​CIC to restore the trust of customers and the market and advance into a global AI company.”

………………………………………………………………………………………………………………………………………………………………………………………………………

References:

SKT, ‘AI CIC’ 출범해 AI 골든타임 잡는다

https://www.businesskorea.co.kr/news/articleView.html?idxno=253124

https://www.lightreading.com/ai-machine-learning/skt-consolidates-ai-capabilities-under-new-business-unit

SKT-Samsung Electronics to Optimize 5G Base Station Performance using AI

SK Telecom unveils plans for AI Infrastructure at SK AI Summit 2024

SK Telecom (SKT) and Nokia to work on AI assisted “fiber sensing”

SK Telecom and Singtel partner to develop next-generation telco technologies using AI

SK Telecom, DOCOMO, NTT and Nokia develop 6G AI-native air interface

South Korea has 30 million 5G users, but did not meet expectations; KT and SKT AI initiatives

 

Page 5 of 344
1 … 3 4 5 6 7 … 344

Archives

Archives

Recent Posts