Is a MSG company blocking the throat of global AI chips? When talking about the bottlenecks of AI chips, the first names that usually come to mind are these: Nvidia’s GPU, Samsung and SK Hynix’s HBM, and TSMC’s CoWoS advanced packaging. These are indeed very critical production links. But you may not think of it: there is a more hidden choke point, hidden deep in the entire supply chain.
The one holding this node is not a semiconductor giant, but a Japanese company that is known to the public as "selling MSG" - Ajinomoto.
What most people don’t know is that in the semiconductor industry, it has another identity:
A near-exclusive supplier of ABF (Ajinomoto Build-up Film), the most critical insulating material in global AI chip packaging.
According to reports from many industry organizations such as TrendForce, Ajinomoto’s global market share in the field of ABF materials used in GPU and CPU packaging substrates exceeds 95%.
Ajinomoto's condiment Masako chicken stock seasoning, few people know that this condiment company also controls more than 95% of the world's supply of ABF, a key material for AI chip packaging.
Ajinomoto defined ABF as the "de facto standard" in the semiconductor market in its 2023 annual report.
This means that almost every high-performance chip in the world, from Intel's CPU to Nvidia's AI accelerator, has to get the thin insulation film in the middle from this "MSG factory."
a thin film
Determine whether the chip can be used
Let’s talk about the simplest metaphor.
The chip itself is very small, and the circuitry on it is nanoscale. But it needs to communicate with the external circuit board, and the circuits on the circuit board are millimeter-level. From nanometers to millimeters, the difference is six orders of magnitude.
How? Depends on the packaging substrate.
There are many layers of microcircuits on the substrate, and each layer leads the signal from the chip to the motherboard. ABF is the insulating film between these microcircuit layers.
A layer of ABF must be sandwiched between each layer of circuits to prevent signal crosstalk and ensure signal integrity.
You can think of it like the sound insulation between each floor in a tall building. Without it, there would be noise upstairs and downstairs, making the whole building uninhabitable.
Same goes for chips.
Without ABF, high-frequency signals would interfere with each other, and the chip would be a pile of waste silicon.
Schematic diagram of the layered structure of the ABF package. The golden highlighted ABF substrate layer in the middle is the high-density interconnection core of the entire package, responsible for ensuring signal integrity at multi-GHz frequencies.
For traditional PC chips, the substrate requires about a few layers of ABF, and the amount is not large.
But AI chips are different. AI accelerators such as NVIDIA Blackwell and Rubin have much larger package sizes than traditional chips, and the number of substrate layers has also increased dramatically.
According to data disclosed at the Ajinomoto business briefing, the amount of ABF used in high-performance CPU packaging substrates is more than 10 times that of ordinary PC substrates.
Some industry analysts also believe that the actual multiple of AI accelerators may reach 15 to 18 times due to more packaging layers and larger size.
The ABF usage of a chip has skyrocketed by an order of magnitude, but there is only one major supplier in the world.
Needless to say, the seriousness of the problem.
NVIDIA Rubin mass production
Go through Ajinomoto first
The Rubin platform, which will be officially released by Nvidia in 2025, will require a higher level of packaging density.
As chips become larger and larger, packaging becomes more and more complex, and the number of ABF layers required increases accordingly.
Traditional packaging may only require a few layers of ABF, but the packaging of AI accelerators often requires 8 to 16 layers or more.
If the size of Rubin and Rubin Ultra increases further, ABF will become the narrowest choke point in the entire supply chain.
Nvidia CEO Jensen Huang launched a new generation of Rubin chips at the 2026 International Consumer Electronics Show (CES 2026) on January 5. The packaging size of AI accelerators is increasing from generation to generation, and the demand for ABF films has skyrocketed.
Ajinomoto himself knows this.
At the latest business briefing, Ajinomoto stated: AI and HPC are driving up demand for ABF, and Ajinomoto promises stable supply.
But commitment is one thing, capacity is another.
According to TrendForce, Ajinomoto plans to invest at least 25 billion yen (approximately RMB 1.2 billion) by 2030 to increase ABF production capacity by 50%.
50% sounds like a lot.
However, compared with the annual double-digit growth rate of AI computing power demand, whether this expansion pace is sufficient is a huge question mark.
Even more troublesome are the technical risks of the expansion itself.
ABF's production process is extremely precise, and yield is the core bottleneck. The more layers there are, the more problems there are in any one layer, which could lead to the failure of the entire multi-layer structure.
Although new processes such as semi-additive patterning (SAP) can improve performance, yield risks also increase.
Ajinomoto ABF film roll material. The ABF film is pressed into the packaging substrate layer by layer and acts as an insulating layer between microcircuits. It is this inconspicuous translucent film that has blocked the throat of AI chips around the world.
This means that Ajinomoto does not want to expand production, but the speed of production expansion is naturally restricted by process yield.
TSMC's CoWoS production capacity is tight, the AI chip delivery cycle is lengthening, and ABF supply constraints are one of the reasons behind this.
Throughout the entire chain, GPU has no shortage of design and HBM has no shortage of production lines, but in the end they are all stuck on a layer of thin film material.
Hyperscale cloud service providers are already aware of this problem.
According to industry reports, some technology giants have begun to help Ajinomoto build new production lines through sky-high advance payments and lock in long-term supply contracts.
When the world's richest company starts paying a deposit for a monosodium glutamate factory to produce capacity, the image itself speaks for itself.
From MSG to chips
Ajinomoto's Invisible Empire
When it comes to this, many people's first reaction is: Why does an MSG company go to make chip materials?
I suspect it is trying to capitalize on the popularity of AI, but in fact it is just the opposite:
Ajinomoto itself is an underrated materials giant.
Ajinomoto was founded in 1909 and started with MSG.
But as early as the 1970s, it began researching the application of amino acid chemistry in epoxy resins and composite materials.
In 1996, a CPU manufacturer approached Ajinomoto, hoping to use its amino acid technology to develop new thin-film insulation materials.
Ajinomoto assembled a team and completed the research and development of ABF in just four months.
In 1999, ABF was officially put into production, and Intel was the first customer.
In the following decades, Ajinomoto maintained a silent monopoly in the ABF field.
In the PC era, the mobile era, and the cloud computing era, this film has been silently lying in the packaging of almost every high-performance chip in the world, but no one paid attention to it.
Until the demand for AI computing power begins to explode exponentially.
Ajinomoto President Taro Fujie mentioned in an interview with Newsweek that ABF's share in the global semiconductor insulation film field exceeds 95%.
People who are reading this article are probably already using devices equipped with ABF, but they may not know it.
Therefore, this is not an MSG company taking advantage of the popularity of semiconductors, but a hidden champion of fine chemicals whose true strength is obscured by consumer product labels.
Every time you use AI
Everyone is paying for this film
Bringing it back to the questions everyone cares about:
Why are AI services so expensive?
Why are Nvidia chips always nervous?
Why are cloud service providers spending so much money building data centers?
Why are the API call fees of Claude, GPT, and Gemini falling so slowly...
There is certainly more than one answer, but ABF is one of the variables that is severely underestimated.
The logical chain is very straightforward:
ABF production capacity is limited, and advanced packaging production capacity is limited; packaging cannot keep up, and AI chip shipments cannot keep up with demand; there are not enough chips, and computing power is in short supply; computing power is in short supply, and services are expensive.
Every time you call a large model, every time you generate a picture, every time you ask AI to write a piece of code for you, the cost structure has the shadow of Ajinomoto's film.
When everyone discusses "AI infrastructure is too expensive," they often focus on the unit price of GPUs, data center electricity bills, and cooling system costs.
But few people realize that the production capacity ceiling of an insulating film material is transmitting pressure upward from the deepest part of the supply chain, and is ultimately reflected in the cost of use of each end user.
The real battlefield for AI competition
Has descended into the periodic table of elements
GPU architecture can catch up. Transformer can be open source and the training framework can be copied.
But chemistry cannot be copied.
Ajinomoto relies not on spending money to build a factory to make ABF, but on the synthesis process accumulated over more than 100 years of amino acid chemistry.
This kind of barrier cannot be solved by investment cycle, cannot be copied by a few engineers, or even can be cracked by reverse engineering.
When AI competition sinks from the software layer to the chip layer, and then from the chip layer to the material layer, the real moat is no longer in the code, but may be hidden in the molecular formula.
This is reminiscent of a recurring script in the semiconductor industry: every round of computing power leaps will expose the weakest link in the supply chain.
The last round was lithography machines, and ASML became the global focus. This time around, the spotlight is turning to packaging materials.
ASML NXE:3400B EUV lithography machine, a single unit sells for more than US$200 million
Ten years ago, no one would associate a MSG factory with AI computing power.
But today, the world's top technology companies are also lining up to sign long-term contracts with Ajinomoto and pay deposits for production capacity in advance.
A hidden bottleneck of computing power is actually hidden in a chemical production line.