CPU, QUIETLY BACK TO THE CENTER OF THE A.I.C.C.U

ORIGINAL TITLE: CPU, QUIETLY RETURNING TO THE CENTER OF THE AI CALCULATOR STAGE
Original by Xu Shan, Optimus Park

IN THE LAST THREE YEARS, THE AI CALCULUS STORY HAS BEEN ALMOST AROUND GPU。

FROM H100, H200, TO GB200, TO GB300, TO CLOUD MANUFACTURERS TAKING THE EXPANDED 100,000-CALORIES CLUSTER -- ALL THE INDUSTRY NARRATIVES SAY ONE THING:THE MATH BOTTLENECK IS IN GPUI DON'T KNOW. CPU, IN THIS STORY, IS A LESS IMPORTANT "COMPATIBLE" ROLE THAT FOLLOWS THE GPU AND IS RESPONSIBLE FOR SOME JOBS THAT GPU DOESN'T WANT TO DO。

But starting in 2026, there was some crack in this narrative。

On June 1st, Intel launched the Power 6+ Processor in Beijing, which is dedicated to clouds, intelligent AI and web-intensive work loads. This is the first CPU data centre for the Intel 18A program。

INTEL'S OWN DESCRIPTION, THE POWER 6+ CARRIES NOT THE GPU'S 'COMPASS', BUT THE AI INFRASTRUCTURE'We have to get to the bottom of thisResponsible for organization, distribution and data flow。

"AI's extension lies not in the overlay of components, but in the coordinated operation of the system. Kevork Kechichian, Executive Vice-President of Intel Corporation and Managing Director of Data Centre Business, said at the conference, "As AI moves towards the age of smarts, the organization and distribution of data flows have become new constraints。

THIS REINFORCES THE CORE FACT THAT THE CPU REMAINS THE CONTROL PLANE OF THE MODERN AI INFRASTRUCTURE

It's not just the Intel family's judgment. In February of this year, SemiAnalysis, an independent semiconductor research institute, published a report on the 2026 data centre CPU map called "CPU Return", which is equally straightforward。AT THE MOMENT, WHEN AI TRAINING AND REASONING HAVE SPREAD ON A LARGE SCALE, CPU IS BEING RE-REQUIRED IN A DIFFERENT WAY FROM THE LAST THREE YEARSI don't know。

IT'S JUST THIS "RETURN" THAT NEEDS TO BE OPENED TO SEE THAT IT'S NOT THE CPU BEING THE LEAD, BUT THE CPU BEING REDEFINED IN A NEW POSITION。

I. THE GPU CENTRAL FRAGMENTATION

TO UNDERSTAND WHY CPU IS "BACK," YOU HAVE TO GO BACK TO THE CHANGES THAT ARE TAKING PLACE IN THE AI WORKLOAD ITSELF。

OVER THE PAST TWO YEARS, THE MAINSTREAM NARRATIVE OF AI ARITHMETIC HAS BEEN TRAINING, WITH LARGE MODELS GROWING FOUR TO TEN TIMES THE SIZE OF THE TRAINING EACH YEAR, AND TRAINING REQUIRES A PARALLEL CALCULATION OF THE SIZE OF THE MASS. BUT TRAINING IS NOT ALL OF THE AI WORKLOAD。

ACCORDING TO INTEL AT THE CONFERENCE, THE OVERALL AI CALCULATOR LOAD CAN BE BROADLY DIVIDED INTO THREE CATEGORIES:

The first one is..Base job load。Storage, database, Web, micro-service, CDN, these are not AI, but AI runs the bottom line. This part remains the traditional CPU's main battlefield。

The second is..Training。THE FRONT-LINE LARGE MODEL TRAINING IS ALMOST ENTIRELY BASED ON GPU AND SPECIALIZED ACCELERATORS. THIS IS THE PART OF THE FIGHT THAT HAS BEEN FOUGHT BY EVERYONE FOR THE PAST THREE YEARS。

The third is..Decomposition with intelligence。This part is growing rapidlyAnd it's quite different from trainingI don't know。

The key difference in the third category is the pattern of the load itself. The training is a process of "calculating" models from nothing to nothing, and it is extremely parallel, and the need for single-point peak computing is extremely high. But reasoning and intelligence aren't -- it's going to be a model that's already been trainedWe'll be in real businessI don't know。

That means a lot of things aren't countingOrganizationmovement of multiple model collaborations, context management, coordination of data flows between different angents, processing of simultaneous requests from users, assurance of delay predictability。

THESE THINGS, GPU IS NOT GOOD AT IT。

"In this scenario, we can see the acceleration of the GPU level, but the subject is still a work load centred on the traditional CPU. Kevork Kechichian spoke at the conference。

There is a more specific industrial fact behind this. SemiAnalysis gave an example in his report "CPU Return": Microsoft built the "Fairwater" data centre for OpenAI, which has been set up by the United States of AmericaA 48 MW CPU AND STORAGE BUILDING THAT SUPPORTS A 295 MW GPU CLUSTERI don't know。

IN OTHER WORDS, IN ORDER FOR THAT 295 MW GPU CLUSTER TO REALLY RUN, IT WOULD TAKE THOUSANDS OF CPUS TO PROCESS THE PB-CLASS DATA FLOWS GENERATED BY GPU, TO SCHEDULE TASKS, TO MANAGE STORAGE。

GPU'S CALCULATIONS ARE PUSHED HIGHERThe greater the "neighborhood" needsI DON'T KNOW. AND THE NEED FOR THESE SURROUNDINGS TO CALCULATE EVENTUALLY FALLS ON THE CPU。

IN OTHER WORDS, CPU'S RETURN IS NOT "CPU AGAIN FASTER THAN GPU". IT'S WHEN THE AI CALCULATOR PATTERN GOES FROM "TRAINING A LARGE MODEL" TO " RUNNING THOUSANDS OF INTELLIGENT BODIES."The organization and data flow has become a further bottleneckI DON'T KNOW. GPU CAN'T FIX THIS, CPU CAN。

THIS IS THE OTHER SIDE OF THE AI NARRATIVE OF THE LAST THREE YEARS THAT HAS BEEN IGNORED。

- What's in? Road

The judgement of intel is reflected in the definition of products up to 6+。

The most intuitive number isUp to 288 coresand all energy efficiency (E-core)。

E-core and P-core are the crosses that Intel has made on CPU structures over the past few years. P-core is a performance core that pursues the design objectives of the CPU, a traditional server. E-core is an energy-efficient nuclear with a weaker mononuclear, but small size and low utility, which can fill more of the same chip area。

As strong as 6+ push this fork to the extreme。288 energy efficient nuclear weaponsTHIS MEANS THAT INTEL'S BET ON A CPU IS NOT "HOW FAST EVERY NUCLEAR" BUT "HOW MANY NUCLEARS CAN FIT ON A CPU."。

The logic of this product definition is:SMART BODY, AI'S WORKLOAD IS NOT HOW FAST SINGLE NUCLEAR POWER RUNS, IT'S THE ABILITY TO RUN THOUSANDS OF LIGHT JOBS SIMULTANEOUSLYI don't know. When a server is required to organize hundreds of delegates, process thousands of reasoning requests and maintain tens of thousands of simultaneous connections, the capacity of 288 E-cores to swallow is far more important than the single-nucleus performance of 64 P-cores。

This is an anti-mainstream product definition。OVER THE LAST FEW DECADES, THE MAINSTREAM NARRATIVES OF THE SERVER CPU HAVE BEEN SPELLING NUCLEAR FEATURESThe main frequency is higher, the IPC is stronger, the cache is larger. The E-core route essentially recognizes that the narrative may be over。

But there are a few things that must be seen together。

First of allE-core's route is not unique to IntelI don't know. AMD launched Bergamo in 2023, the Zen 4c core based on density optimization, the Graviton series of AWS, the AmpereOne series of Ampere, and the "High Density Core + Energy Efficiency Priority" route long ago. Ampere has achieved 512 core numbers in the AmbereOne Aurora Road Map published in 2024。

Intel is chasing an industry that already existsIntel is not a leader. He's a player back in industryI don't know。

SecondAs strong as 6+ is the first data centre for the Intel 18A program CPUThis may be more important in Intel’s own language than the Nuclear E-core。

Intel 18A is Intel's biggest bet in the last few years. It's not just a CPU thing, it's an Intel Fundry, Inteldeer, stand-up thing. The story of Intel Foundation cannot go on if the 18A program does not give the market a hit。

TO THE STRENGTH OF 6+ MADE WITH 18A PROGRAMS, ENERGY EFFICIENCY NUMBERS PUSHED TO 288, AND THE "PERFORMANCE DENSITY INDUSTRY LEADS" ANNOUNCED TO THE PUBLICThis is one of Intel's responses to the marketI don't know. It is another question whether it can be recognized by the market and stand in competition with the generation of the N2, Samsung 2nm。

Thirdly, the top 6+ customer list contains several industrial names - Eric to the top 6+, testing 5G core network, T-Systems under the German telecommunications flag, using as much as 6+ private intelligence AI infrastructure. Both clients are the traditional, robust suppliers of CPU data centresTheir procurement choice is itself a market signalI don't know。

I'm not sure what I'm talking aboutThat's the way to betThe energy efficiency advantage is obtained by the 18A process, the core density is obtained by 288 E-cores, and the loads of jobs such as "high density, energy efficiency, high throughput" in the AI reasoning and intelligent body scene。

IT'S NOT THE STORY OF CPU REGAINING POWER, BUT THE STORY OF CPU FINDING A NEW LOCATION。

III. Whether or not this is going to work

DID INTEL TELL YOU ABOUT THE CPU RETURN? WE NEED TO LOOK AT A FEW OTHER VARIABLES IN THE INDUSTRY。

THE FIRST VARIABLE IS THE REACTION OF GPU MANUFACTURERS。

The combination of Grace CPU+Hopper GPU, which is itself a complement to the CPU, is also being done by Yvette over the past two years. If the GPU producers themselves mainstream the overall CPU+GPU program, then the position of CPU producers as independent players will be reduced。THIS IS INTEL'S BIGGEST OPPONENT OF THE STORY OF "CPU IS THE CONTROL PLANE," NOT AMD, IT'S YIN WIDA HIMSELFI don't know。

THE SECOND VARIABLE IS THE SELF-STUDY CPU OF CLOUD MANUFACTURERS。

AWS Graviton has been deployed in its own data centre, taking on a significant portion of the general computing load within AWS. Microsoft is doing Cobalt, Google is doing Axion, Ali is doing trustALMOST ALL MAJOR CLOUD MANUFACTURERS ARE SELF-STUDYING THE ARM ARCHITECTURE SERVER, CPUI don't know。

THESE SELF-STUDY CPUS ALSO FOLLOW THE "HIGH DENSITY, ENERGY EFFICIENCY PRIORITY" ROUTE - AND UP TO 6+ ARE DIRECTLY COMPETITIVE IN PRODUCT DEFINITION。

That is, the top 6+ market that you want to takeThe cloud dealers are doing it themselvesI DON'T KNOW. INTEL NEEDS TO PROVE THAT THERE ARE STILL ENOUGH MARKETS OUTSIDE THE CPU FOR CLOUD MANUFACTURERS. FOR EXAMPLE, TELECOM OPERATORS, PRIVATE CLOUD, AND VERTICAL INDUSTRY DATA CENTRES。

THE THIRD VARIABLE IS THE 18A PROGRAMMING ITSELF。

As strong as 6+ is the first CPU data centre for Intel 18AThis in itself means that the chip has taken on far more industrial significance than the product itselfI don't know. If the 18A process has problems with yield yields, performance stability, customer certification, the strong 6+ market performance will be stalled. On the other hand, if 18A is stable, 6+ might bring some breathing space to Intel Foundation。

But 18A wasn't running in a vacuum -- the N2 process, which accumulated electricity, started production in the second half of 2026, and Samsung 2nm was on the way。Intel 18A wants more than "do it," "do it and lead."It's a higher standard。

Combining these three variables, the final color of the 6+ depends not only on itself, but also on whether the CPU role will be eaten by Young Weidar itself, whether the Cloud Factory Chamber will continue to study CPU, Intel 18A, or whether it will stand in competition with the same generation as the station and the Samsung。

THIS IS WHY THE CPU RETURNEDFrom an industry-level perspective, it's still not known if Intel can get the return dividend himselfI don't know。

CPU'S POSITION ON THE AI CALCULUS STAGE HAS BEEN CONTESTED FOR THREE YEARS。

THE SCRIPT FOR THE LAST THREE YEARS IS "GPU IS THE CENTER, CPU IS THE CENTER." THIS SCRIPT STARTED LOOSE IN 2026 -- NOT CPU AGAIN FASTER THAN GPU, BUT AI'S OWN ALGORITHM IS CHANGING. WHEN AI EXPANDED FROM "TRAIN A MODEL" TO "RUN THOUSANDS OF INTELLIGENT BODIES", ORGANIZING, CO-OPTING, AND DATA FLOW BECAME SYSTEMIC BOTTLENECKS, CPU BECAME IRREPLACEABLE IN THIS POSITION。

Intel put this in, up to 6+ is the reply it gave. But is it possible that Intel will be able to get the dividends himselfIn the end, the answer is in the client rooms of 2027, 2028。AMD, ARM, CLOUDMAKER, CPU, YVETTE, AND CPU, EACH VARIABLE CAN CHANGE THE COURSE OF THE SCRIPT。

THE CPU'S RETURN IS TRUE, BUT WHO'S GOING TO TAKE THE LEAD ON IT。

Original Link