<\/div> Follow ZDNET: Add us as a favorite source<\/span><\/a><\/span> On Google.<\/em><\/p>\n The sudden \u2013 even by technological standards \u2013 shift towards AI has not left businesses scrambling to catch up; Even America’s spy agencies are struggling to keep up.<\/p>\n Too: <\/strong>AI isn’t getting smarter, it’s becoming more power-hungry and expensive<\/strong><\/p>\n This is why the government has given a thumbs up Secret $9 billion request for superchips<\/a> This will help the CIA and NSA keep pace with the likes of big AI players anthropic<\/a> And OpenAI<\/a> Are doing. <\/p>\n Running the latest batch of AI models requires massive amounts of computing power \u2013 not to mention the massive supplies of power and specialized cooling that come as part of cutting-edge, modern data centers, and the silicon to deliver it is Nvidia’s Grace Blackwell superchips, named after the American mathematician. david blackwell<\/a> and American computer scientist and naval pioneer grace hopper<\/a>. <\/p>\n Also: How to learn cloud code for free with Anthropic’s AI courses \u2013 one only took me 20 minutes<\/strong><\/p>\n These superchips are called gb10<\/a>The 20-core Arm CPU, codenamed Grace, manufactured by MediaTek is based on the Blackwell architecture with an Nvidia GPU. Take this chip and add 128GB of LPDDR5x \u2013 it’s the demand for memory for AI that has caused the price of things like RAM and Raspberry Pi boards to skyrocket \u2013 and 4TB of storage in the form of an NVMe M.2 SSD, and you have a chip that delivers 1 petaflop of FP4 AI performance for just 140 watts of power draw. He’s just a chip. <\/p>\n That’s not much when you consider that most modern gaming PCs can swallow up to 1,000 watts of power. <\/p>\n How does the scaling blackwell architecture work?<\/p>\n<\/div>\n<\/div>\n NVIDIA<\/span><\/figcaption><\/figure>\n This one chip has the power to fine-tune AI models with 70 billion parameters. In terms of storage alone, a model like this requires around 140GB of space. <\/p>\n Want a GB10 system? Best Buy sells a rack version Starting around $5,000<\/span><\/a><\/span>! <\/p>\n Too: <\/strong>What Google’s TurboQuant can and can’t do for the rising cost of AI<\/strong><\/p>\n But the real power usage happens when you scale them up. The GB300 NVL72 racks 72 GPUs and 36 CPUs in a single, liquid-cooled unit. Now scale this up to data center proportions, and you can begin to see why power demand is growing exponentially. <\/p>\n The price of a rack can be anywhere in between $1.8 million and $4 million<\/a>. And a data center can have more than 100,000 racks. <\/p>\n But if you want to run larger AI models like Anthropic’s Cloud, OpenAI’s GPT 5.5, or DeepSeek’s v4, this is what you need. <\/p>\n AI is seen as both a next-generation tool and a national security threat, something that is again moving faster than governments can legislate for or put guardrails in place. Just the other day, a planned executive order<\/a> It would have outlined a process where AI companies would “volunteer” their models for government testing for a period of 90 days before public release, which was canceled after pressure from industry leaders. <\/p>\n <\/div> A Dell GB10 desktop AI computer<\/p>\n<\/div>\n<\/div>\n Ditch<\/span><\/figcaption><\/figure>\n The order makes it clear that the government not only wants to leverage AI, but it also wants to be able to scrutinize the models used by the public. <\/p>\n Too: <\/strong>How Qualcomm’s new wearables chipset could end smartphone dominance<\/strong><\/p>\n This will require serious hardware horsepower. <\/p>\n There is also an element of coping with the lack of investment in computing hardware over the years. Combine this with the current shortage of chips and other AI hardware, and it means billions will have to be spent to stay in the game. <\/p>\n The $9 billion, which still needs to be approved by Congress, will allow the government to acquire both the infrastructure and hardware needed to remain relevant in the AI \u200b\u200bgame. <\/p>\n Inside the GB10-based AI system.<\/p>\n<\/div>\n<\/div>\n Ditch<\/span><\/figcaption><\/figure>\n But buying chips and expanding data centers takes time, so in the interim, about $800 million of the defense budget has been repurposed to get more cloud computing power. Intelligence services are also continuing to use an advanced AI model called Anthropic mythology<\/a>Despite the company being described as a threat to the supply chain. <\/p>\n And that $9 billion, although it sounds like a lot, doesn’t really count in the grand scheme of AI. Amazon Web Services is is investing $50 billion to upgrade its government cloud computing services<\/a>A platform that is used extensively by intelligence agencies. <\/p>\n Also: I left ChatGPT for a free, private, and local AI called Olama \u2013 here’s why<\/strong><\/p>\n And Grace Blackwell Silicon has a successor in the pipeline – the vera rubin<\/a> The stage, named after an American astronomer. These chips combine a brand-new, custom-built Arm-based CPU called Vera and a high-performance GPU called Rubin, and are designed to offer up to 10 times more performance per watt than Grace Blackwell and use high-performance GPUs. HBM4 memory<\/a>. <\/p>\n AI has now become a modern arms race, and governments that want to keep up will have to invest a lot of money. <\/p>\n<\/div>\n ZDNET Highlights <\/h3>\n
\n
So what are these superchips? <\/h2>\n
<\/picture><\/div>\n<\/p><\/div>
Why does the government need so much power? <\/h2>\n
<\/picture><\/div>\n
<\/picture><\/div>\n<\/p><\/div>
$9 billion is just the tip of the iceberg <\/h2>\n