Memory is emerging as the biggest bottleneck in AI data centers, even as investments in processors, cooling, and rack density have accelerated. ScaleFlux is developing controller chips targeting the external DRAM and SSD stages of the AI data pipeline, aiming to address capacity and bandwidth constraints as workloads shift toward inference and agentic tasks.
Microsoft is exploring composable memory architectures that pool memory beyond individual servers and assign it to workloads on demand. The company says interconnects, including optical links between memory and processors, are critical to making composable memory viable at scale, and advises enterprises to prioritize their top workloads when building vertically integrated systems.
AMD took the stage at the OCP Summit to unveil its Helion server rack, built on standardized Open Compute Project specifications. The rack delivers modular computing capacity and supports agentic AI workloads alongside industry partners, targeting hyperscale facilities that require gigawatt-scale power delivery while reducing deployment delays and vendor lock-in.
On the training side, the sheer scale of compute needed for AI models is staggering. OpenAI's GPT 1 required 18 million trillion calculations in 2018, while GPT 6 Astra in 2026 took nearly an octillion, over 50 million times more. Researchers expect training demands to keep climbing through 2030 as scaling laws continue to push compute requirements upward.
Check Point executive chairman Gil Shwed, who co-founded the company in 1993, said on Calcalist's Money Engines podcast that he does not see economic models justifying the hundreds of billions flowing into AI infrastructure. He called AI potentially the greatest technological revolution yet but stressed that judgment remains premature.
Meanwhile, US and China are reportedly working to make AI cooperation a deliverable when presidents Xi Jinping and Donald Trump meet next month. Technical teams from both countries are shaping the AI agenda ahead of talks expected on the sidelines of a summit in Shenzhen.
Congressman Josh Gottheimer visited the University of Pennsylvania on October 12, 2026, to discuss federal AI policy. He co-chairs the House Democratic Commission on AI and highlighted bipartisan bills including the American AI Security Act, Stop Rogue AI Act, AI LABS Act, and Liquid Cooling for AI Act. He also praised Penn for launching the Ivy League's first undergraduate AI major.
Quality concerns have surfaced around OpenAI's release of over 300 math proofs. Mathematician Sophie Huiberts, writing on Terence Tao's blog, called the proofs poor quality, citing lack of clear explanations, errors that could warrant retraction, and missing code. Tao amplified these concerns by hosting the post.
In healthcare, a new FDA-cleared AI tool called TumorSight Risk, developed by SimBioSys with NVIDIA technology, predicts breast cancer recurrence risk from existing MRI scans. A study of 2,129 women showed high-risk patients faced more than three times the risk of cancer returning compared to low-risk patients, with scores ranging from zero to 100.
On the hardware cost front, shifting AI workloads to local systems can cut long-term compute expenses by up to 60 percent. Local setups with GPUs, fast RAM, and NVMe storage also reduce cloud subscription costs, bandwidth delays, and security risks for sensitive data.
A panel in Fayetteville, Arkansas addressed the growing misuse of AI to generate explicit images of children. Leah Juliett, whose intimate images were shared without consent as a teenager, spoke about the harm caused when AI enables someone to transform a child's photograph into illegal explicit content.
Key Takeaways
- ScaleFlux is building controller chips to address memory and storage bottlenecks in AI data centers as inference and agentic workloads grow.
- Microsoft is pursuing composable memory architectures, emphasizing optical interconnects and workload-specific system design to tackle AI memory scarcity.
- AMD unveiled its Helion Rack at the OCP Summit, using Open Compute Project standards to support agentic AI in hyperscale environments.
- OpenAI's GPT 1 required 18 million trillion calculations in 2018, while GPT 6 Astra took nearly an octillion in 2026, reflecting exponential training growth.
- Check Point's Gil Shwed says no economic model yet justifies the hundreds of billions invested in AI infrastructure.
- The US and China are reportedly negotiating AI cooperation ahead of a potential Xi-Trump meeting next month.
- Congressman Gottheimer promoted bipartisan AI bills including the American AI Security Act and Liquid Cooling for AI Act at the University of Pennsylvania.
- Mathematician Sophie Huiberts criticized OpenAI's 300+ math proofs as poor quality, with errors and missing code, on Terence Tao's blog.
- FDA-cleared TumorSight Risk, built with NVIDIA tech, predicts breast cancer recurrence from existing MRIs, with high-risk patients facing over 3x the return risk.
- Local AI hardware can reduce long-term compute costs by up to 60 percent compared to cloud processing.
ScaleFlux targets AI memory bottleneck with controller chips
AI data centers have invested heavily in faster processors, denser racks and liquid cooling, but the data pipeline feeding those processors has not kept up. ScaleFlux is building controller chips for the external DRAM and SSD stages of the AI data pipeline. As AI workloads shift toward inference and agentic tasks, memory and storage capacity and bandwidth are becoming the main infrastructure bottleneck.
Microsoft explores composable memory to ease AI supply crunch
Memory is now the scarcest resource in AI data centers, and hyperscalers are rethinking how to allocate it. Composable memory pools capacity beyond individual servers and assigns it to workloads on demand. Microsoft says interconnects between memory and processors, including optical connections, are critical for making composable memory work at scale. The company advises enterprises to focus on their top workloads and build vertically integrated systems around them.
US and China race to shape AI agenda before Xi-Trump meeting
The United States and China are reportedly in discussions to make AI cooperation a potential deliverable when presidents Xi Jinping and Donald Trump meet next month. The technical teams of both countries are working on shaping the AI agenda ahead of the talks. The meeting is expected to take place on the sidelines of an upcoming summit in Shenzhen. Both sides aim to position cooperation on artificial intelligence as a concrete outcome.
Mathematicians criticize OpenAI's 300 math proofs as poor quality
Mathematician Sophie Huiberts wrote a guest post on Terence Tao's blog calling OpenAI's 300+ math proofs poor quality work and easy rejects. Huiberts argues the proofs lack clear explanations, proper attribution and enough context for other researchers to build on them. She says the release includes errors that could warrant retraction and sometimes promises code that is not supplied. Tao has amplified these concerns by hosting the post on his blog.
Check Point's Gil Shwed questions AI investment economics
Gil Shwed, executive chairman of Check Point, says he does not see economic models that justify the hundreds of billions of dollars flowing into AI infrastructure. Shwed spoke on the 400th episode of Calcalist's Money Engines podcast. He called AI potentially the greatest technological revolution yet but said the judgment is premature. Shwed co-founded Check Point in 1993 and pioneered the modern firewall.
FDA-cleared AI tool TumorSight Risk predicts breast cancer recurrence
A new FDA-cleared AI tool called TumorSight Risk can predict if breast cancer will return. Developed by SimBioSys with NVIDIA technology, it analyzes a patient's existing breast MRI to estimate recurrence risk over five to ten years. A study of 2,129 women showed high-risk patients had more than three times the risk of cancer returning compared to low-risk patients. The tool gives a score from zero to 100 and does not require any additional scans beyond what patients already had for diagnosis.
Gottheimer discusses AI policy and bipartisan bills at University of Pennsylvania
Congressman Josh Gottheimer visited the University of Pennsylvania on October 12, 2026 to discuss federal AI policy. He co-chairs the House Democratic Commission on AI with Representatives Ted Lieu and Valerie Foushee. Gottheimer highlighted several bipartisan bills including the American AI Security Act, Stop Rogue AI Act, AI LABS Act, and Liquid Cooling for AI Act. He also praised Penn for launching the Ivy League's first undergraduate AI major.
Training AI models requires mind-boggling amounts of math and computing power
Training artificial intelligence models requires an enormous number of calculations that grow exponentially over time. OpenAI's GPT 1 took 18 million trillion calculations in 2018, while GPT 6 Astra in 2026 took nearly an octillion calculations, over 50 million times more. These figures are estimates because leading AI labs do not disclose specific training details. Researchers predict that by 2030, the power needed to train the largest AI models will continue growing rapidly due to a pattern known as scaling laws.
Northwest Arkansas panel discusses AI-generated explicit images of children
A panel discussion in Fayetteville, Arkansas focused on how artificial intelligence is being used to create and share sexually explicit images of children. Leah Juliett, whose intimate images were shared online without her consent as a teenager, spoke at the event about the harm this causes. She expressed concern that AI makes it possible for someone to take a photograph of a child and create illegal explicit images from it.
AMD unveils Helion Rack and open Agentic AI hardware at OCP Summit
AMD showcased its new Helion server rack at the OCP Summit, built using standardized Open Compute Project specifications. The rack uses AMD processors to deliver modular computing capacity for data centers. AMD displayed platforms built with industry partners to handle complex agentic AI workloads in standard installations and hyperscale facilities requiring gigawatt scale power delivery. The open architecture aims to reduce deployment delays and prevent vendor lock-in for operators.
Local hardware cuts AI costs and cloud risks
Shifting AI workloads to local hardware reduces long-term compute expenses by up to 60 percent. Cloud processing creates high subscription costs, bandwidth delays, and security risks for sensitive data. Workstations need GPUs with large VRAM, fast RAM, and NVMe storage to handle large models. Local systems also support multi-GPU setups and better data privacy for creative teams.
Sources
- Liquid cooling aside, AI memory and storage are the bottleneck
- Composable memory helps Microsoft tackle the AI memory crunch
- Exclusive | US, China race to shape AI agenda ahead of Xi-Trump talks in Shenzhen: sources
- OpenAI’s 300+ Proofs Are “Quality Of Work”, “Easy Rejects”: Mathematician Sophie Huiberts In A Post Shared By Terence Tao
- Gil Shwed: “I don’t see the economic models that justify AI investments on this scale”
- New FDA-cleared AI tool could help predict if breast cancer will return
- RELEASE: Gottheimer Discusses AI Policy at the University of Pennsylvania
- What Are Data Centers Doing? A Mind-Boggling Amount of Math.
- Northwest Arkansas panel discussion focuses on how AI is being used to make, share explicit images of children
- AMD Showcases Helion Rack and Open Agentic AI Hardware at OCP Summit
- High-performance hardware requirements are shifting rapidly as local artificial intelligence tools expand
Comments
Please log in to post a comment.