Stop building the Platforms AI | Looking at the data science

And middle companies benefit success in creating data and ML platforms, creating AI platforms now has been a major challenge. This post discusses three important reasons why you should be careful about building AI platforms and proposed my thoughts on promising directions.

And a statement: Based on personal looks and does not apply to cloud suppliers and data / ml companies. They should also sit down in the use of the Ai platforms.

Where I'm from

In my previous article From Data Platform to ML Platform At times the database science, I did not know how the data platform appears on the ML platform. The trip is working very small and medium-sized companies. However, there is no clear way for small and middle-sized companies to continue developing their platforms on AI forms now. Measures up on AI platforms, the way in which two indicators:

• AI infrastructure: “new electricity” (AI Im Impression) works well when it is more productive. It is a great provider's sport and high model providers.
• The Applications Platform: Unable to build “beach house” (AI platform) in regular surfing. The ability to appear in AI and express new paradigms that make it possible for challenging the normal challenge.

However, there are still indicators that they may be important as AI models continue to appear. Covered at the end of this post.

A higher AI infrastructure

While the databicks maybe only several more spark activities, Deepseek can be 100x functioning more efficient than you find the LLM inferring. Training and service of the LLM model requires the most important investment in infrastructure and, which is important, control of the llM model shape.

In this series, we summoned LLM training infrastructure, including strategies for the same training, toopology design, and speed. On the Hardware side, outside the GPU Performement Performance and TPUS, an important portion of the costs attended the setup network and service delivery services. Collections require additional RDMA network to enable unrestricted connections, identify links to data exchange between conditions. Organic Services must support the complex function of work, Fail Wover plans, the acquisition of hardware problems, and the release of GPU services and entertainment. The training SDK needs to facilitate asynchronous view, process data, and the quality of the model.

In terms of model servitude, model suppliers often include efficiency in the internalization of model development. Model suppliers may have better balance strategies, producing similar model quality in size of a very small model. Model suppliers may develop a better compatible strategy because of the control they have more than the model structure. It can increase batch size during the detection of the acquisition, which effectively increases the use of GPU. Additionally, senior llm players have visual benefits that enable them to reach cheap routers, mainframes, and GPU chips. Most importantly, controlling the powerful model shape and a better model model such as it can mean the power of cheaper GPU devices. For examples who rely on open models, GPU decrease can be a great concern.

Take Deepseek R1 as an example. Suppose you use an example of P5e.48xlage AWS to provide 8 H200 chips with connected Nvlink. It will cost you 35 $ per hour. Thinking about what you do with NVIria and achieve 151 / second tokens. Producing 1 million mile tokens, it will cost you $ 64 (1 million / (151 * 3600) * $ 35). How much Deepseek selling its token in a million? 2 $ only! Deepseek can reach 60 times up to the shipping of your cloud shipping (takes 50% from Deepseek).

Therefore, the power of the LLM feed is like electricity. Indicate variation of the llms requests can fall power; It also means that it works very well when it is produced in the middle. However, you should still handle llm's llm officers for privacy, such as hospitals with its emergency electricity.

Land Change of Land

AI infrastructure investment is a brave game, and build a heavy platforms of AI to come with its hidden issues. By the immediate appearance of AI Model models, no paradigm aligned for Ai application; Therefore, there is a solid basis for building AI programs.

The easier answer is that: Be patient.

After several years of an aggressive competition, few of the biggest characters always stand on the field. However, AI's ability to skill has not yet changed. With the development of AI models, existing development paradigm will soon end. Big players have already started stabbing her agent development platforms, and new solutions appear as popcorn in oven. The last winners will appear, I believe. In the meantime, the agents of building agencies itself is a tricky and medium telemen.

Relief of the old success of the success

Another challenge to build AI platform is hiding. It is about showing a platiform builder, either dependent on the way from previous building data and ML platforms.

Since we had previously shared, from 2017, data and ML Development Paradigms is well-aligned, and the most important work of the ML platform for distribution and issuance platform. However, the Paradigm of the Development of AI application has not been established. If the party follows the story of the front success of the data and ML, it may eventually prioritize the time. The possible guidelines are:

• Build a AI model Gate: provide medical books and applications in the LLM models.
• Create a framework for Agen Agent: Developing a SDK-designed SDK by creating AIs Agents for advanced communication in an internal ecosystem.
• Make normal rag practices

Those programs may be important. But ROI is really dependent on your company scale. Even if you will have the following challenges:

• Continue the recent AI advances.
• Customer acquisition rates where customers are easy for your release.

Suppose the ML Builders and ML platforms are like “the editors of the beds”, AI builders are now they should do as “fashion designers”. It requires accepting new ideas, conducting immediate tests, and even accepts the level of imperfection.

My thoughts in the direction of promising

Whether so many challenges are now coming, please be remembered that it is still satisfying the AI ​​platform now, as a large lonely held:

• Ai-free alternative power is more than data and machine learning.

• The motivation of using AI is more powerful than before.

If you choose to direct the right and plan, the conversion you can bring to your organization is important. Here are some of my thoughts in tracks that can get a little disturbance as AI model scales continue. I think they are equally important about Ai Platistation:

• Modern, Rich Medical Data Products
• Multi-Moderal Data Act: OltP, Olap, and SQL, and Elastics Search, Formal Information Service After the MCP server, may require many types of data supporting high performance data. It is a challenge to maintain a single true source and work with the evolved ETL activities.
• AI Developer: Ai-Centron Software Development, Records, and Analytics. The accuracy of the gene is highly increased in the last 12 months.
• Testing and monitoring: Added uncertainty of AI applications, the assessment and monitoring of these applications is very important.

These are my thoughts in creating ai platforms. Please let me know your thoughts about it. Cheers!