Now, let's move on to our next speaker, Ms. Rosalina Hiu, the Global Vice President of Brand Strategy at Seagate, a leader in hard drive innovation. She will discuss demands of the AI era. Welcome, Ms. Hiu. The floor is yours.
I'm so happy to be here in Computex, and I'm Rosalina Hiu. I manage the global brand strategy for Seagate Technology, and I'm here to share with you how AI storage, specifically hard drive, shape the AI future. And before I do that, because I can, I wanna actually say this: I wanna celebrate all the female in the tech industry in this room. So give it applause. Keep growing and keep believing in your career. So, you and I are living in an amazing time. This is the time of the biggest disruption, potentially, in the humankind history, and this kind of a moment had happened before in the past. It's comparable to, if you look at the history of invention, steam engine, accelerating the global industrialization.
The personal computer, the internet, the smartphone, it accelerated information sharing and gathering, and also share how we interact with each other, how we learn, and how we do business. Gen AI potentially can do that big of transformation or even bigger. It will change the fabric of our life. Now, with the past innovation, it has connection to data, but Gen AI is rooted in lots and lots of data. You heard from the other speakers about trillions of token, parameter, et cetera. Those are data, and data fuel AI, and AI requires more data to make it better. AI also will fuel data generation, new data generation.
This new data generated by AI, the Gen AI specifically, will needs to be grounded with that original data that are being trained as the source of truth, meaning the data really becomes so valuable than ever in this AI era. Storage is super critical to preserve the data for AI. According to the IDC, the global known analyst, they're predicting that by 2028, there's gonna be about 394 zettabytes of data. It's a lot of data, and they already included some prediction of Gen AI in here. It essentially double in size in five years. This forecast, however, is actually understated because it does not include yet the potential of multimodality LLM or what we typically call also the vision model, because it's really hard to predict.
We're still in the very beginning of this vision model, and as it goes ubiquitous, the growth rate may become more than what we expected. Gen AI is truly a data creation force multiplier. The prediction from Gen AI alone, it will create about 100 zettabytes of data in four years. So if you remember the number in the previous slide, that's approximately about 25%, but again, it's understated. And some sources say it's about 170%-200% CAGR for Gen AI to generate data. That's actually significantly more than when the era of smartphone and PC in terms of data generation. And why is it that big? Number one, it's that richer content. With the multimodality Gen AI that involves image and video, no surprise, the data size is huge.
So then most likely, the data that is gonna be replicated also gonna get bigger. And we need this replication because in order for all this new generated data from image and video to be used, it needs to get closer and closer to wherever the use cases are or the users are. So more replication is gonna be needed. And additionally, we believe that more retention is needed for the data. I mentioned about the source of truth. If you hear or read some of the recent news, AI can hallucinate. It's not something that we need to be negative about, it's just something that we need to go through as the technology getting mature. But it is concerning if there's no source of truth.
One of the funny example that I remember is the Gemini, when it's being asked about who are the Founding Fathers of America, it was highly, highly more biased towards diversity answers, which is completely wrong. And if we don't preserve our historical information, then most likely, the distortion of truth can happen years, years after. So it is critical to actually retain source of truth of data. And the governments all over the countries and the world requiring also data retention policy. So we believe with Gen AI, that retention policy gonna get stronger, and it could be indefinitely. So the conclusion here is that it's just gonna be more exabytes that we have to deal with in the world. So for every technology that are introduced in the human mankind, there is the adoption that's called the S-curve.
GenAI is actually following this S-curve too, and it actually accelerates. We believe that when it gets to the mass adoption, the critical thing that everyone has to be prepared for is that scale storage demand. Now, there are technically four stages here, and let me just go through quickly the stages. The first two stages, essentially, where today we are now, it's about building and training the model, and then also deploying the model and creating application on top of these models. It's a lot of excitement. A lot of the startups are actually proliferating in every country, trying to build new application. Then after that, is the early adoption of those application, and then the mass adoption is gonna happen, and this is where the storage is gonna take off.
But in each of these phases, technically even the first two phases, storage are critical too. And you can think about all the data that's needed, like I mentioned, to train, it require that data lake to support it, and most likely it's using the existing capacity of storage. However, with the vision model, we believe that more and more of this data lake gonna have to grow, even the existing capacity will not be enough. For AI to reach full potential for consumer and businesses, it does need to be deployed everywhere: at the cloud or the core, at the edges, which is typically on-prem or colocation, and also at the endpoint, where the consumers are using the application.
So cloud is where the foundational large language model are being stored and trained, but you also heard lots of the announcement from Intel, AMD, and NVIDIA about NPU, that will enable compute for AI PC and AI workstation, and also Qualcomm Snapdragon for mobile AI compute. With all this multimodality, large language model development, the data size is gonna get bigger, and that means also in each of these deployment state, we will need mass storage capacity at the cloud, the edge, and the endpoint. So as AI going everywhere, storage will also be everywhere. So it is not enough for company or any company that are data-intensive to think only about compute and networking requirement to be AI-ready. Storage is the backbone for data preservation that is needed for AI.
Existing storage capacity will not be sufficient, so enterprise will need to be strategic to plan ahead for that storage growth as the mass deployment gonna happen. So data centers must quickly also increase storage capacity, and they have to balance these two forces going against each other, the explosive data creation and resource scarcity. The explosive data creation competes against the resource capacity constantly. Let me go through the resource scarcity that I mentioned here. Real estate. I think you heard from everyone already how data centers are growing like crazy. It's the hottest asset because of AI. I read an article that say that the availability of colocation and data center in Singapore, Tokyo, and Hong Kong is below 2%. All the companies are growing their data centers and looking for land, and it jack up the price of real estate.
The CAGR of data centers globally is about 10%. However, the hyperscalers grow at 20% rate. So everyone is competing. The colocation space, essentially, is very competitive in terms of prices, and the cost to build data center is also growing up to $1.5 billion. Second is the power. It is very energy hungry. I think everyone before me already mentioned this as well. It's about 160% more power demand AI is gonna need by 2030. And considering one GenAI image is being generated, it takes about the same as the energy of one fully charged smartphone. So yes, we do need to figure out how this power issue is gonna be. Now, when it comes to storage decision, besides the sustainability aspect, is the budget.
There is a TCO, the terms of total cost optimization for storage decision process, and that is acquisition of the devices, the power, and other costs. But the largest cost out of the three is actually acquisition of the devices. So how do enterprises and hyperscalers decide on this? Now, let's use hyperscalers, because they are the poster child as an example of the best in terms of TCO optimization. So I'll let you know the secret. It is not SSD that they use the most. It's actually 90% of their data they store in hard drive. And why is that? Because of the TCO of the acquisition cost of hard drive. It's 6-to-1. It will forever be 6-to-1 ratio for SSD for the foreseeable future.
So given that, hard drive is a fundamental component for cloud storage and any data center that needs to scale out for GenAI. With that, we do need to innovate and figure out how to support them. So hard drive areal density is the key to support the key challenges of AI, data center, and AI storage. How do we increase the capacity of hard drive? Technically, there are two ways here. The easy way to think about it is that in the same form factor, with the current PMR, or perpendicular magnetic recording, you can add more platter. At the maximum right now, there's only 10 platter or 10 disks inside the hard drive. You can add 11th platter to increase the capacity.
It sounds simple; however, it actually adds more cost to the materials, and it also increases the operating costs, power consumption, as well as resource usage. The other way is the most sophisticated one, and not easy to get to, is to actually pack more bits in the same materials of the form factor that we have. This is the most sustainable solution, and it actually helped the company to lower the power consumption as well. So Seagate, in early this year, had introduced Mozaic 3+ platform. This is the highest areal density you can achieve in hard drive industry right now. The areal density that we have achieved is 3 TB per disk, which is far more compared to the past platform, which is around 2-2.4 TB per disk.
Mozaic 3+ is a composite of the most complex nanoscale recording technologies and material science breakthroughs on the industry. It is as complicated or even more with the semiconductor industry. It is our decades of R&D to achieve this density, and is extensible beyond 3 TB per disk. More bits are packed in the same form factor, 3.5-inch hard drive, no additional platter or head. It's compatible with existing system configuration, so it's easy to deploy and transition for any data centers. So what is the value of areal density increase? It's simply scale, again, TCO optimization and sustainability. And let me quickly go through that. The impact of areal density at scale is profound.
Think about this example: If a data center, in average, they have a fleet of 16 terabyte in their data centers racks, and they upgraded to 30 terabyte, which is basically our 3 TB per disk platform, Mozaic 3+, it essentially delivers double the capacity of storage in the same data center with the same floor space. And what about the power? With the same scenario from 16 terabyte upgrade to 30 terabyte, they will see at least 45%+ power savings. You can see the difference of the number here. The 16 terabyte consume about 0.59 watt per terabyte, whereas the Mozaic 3+ consume about 0.32 watt per terabyte. So roughly, it's about 45%. And we also observe greater than 55% embodied carbon reduction, and it's great for the sustainability aspect.
So it's about storing more without increasing consumption of space, power, and natural resources. Additionally, the new platform, it was made with 28% less of recycled material by weight, and that also reduces the packaging and shipping impact. Areal density is truly an enabler for a sustainable datasphere and circular economy. The areal density breakthrough is not just what you see with that, what we introduced, Mozaic 3+. Tomorrow's advancement is already here now. In our roadmap, in our lab, we already achieved the 4 TB per disk and 5 TB per disk density. We have the prototypes, and we believe that we can deliver 50 TB drives by 2028. So that's about doubling the capacity for every four years. In the world of AI, data upholds AI, and storage, hard drive, upholds data.
Hard drive areal density technology breakthrough is super critical for AI storage to scale and to be sustained. It is where we believe the future of AI is read and written. Thank you, and we hope to see you in fourth floor.
Thank you, Miss Hiu, for showing us where the future is being read and written.