---
title: "Smoltek Semi & AI"
canonical_url: "https://www.smoltek.com/smoltek-semi-ai/13551/"
date: 2025-12-17
author: "Fredrik Liljeberg"
featured_image: "https://www.smoltek.com/wp-content/uploads/2025/09/smoltek-ai-society-1920x1080-1.webp"
categories:
  - name: "IR Blog Posts"
    url: "https://www.smoltek.com/category/ir-blog-posts.md"
tags:
  - name: "advanced chips"
    url: "https://www.smoltek.com/topic/advanced-chips.md"
  - name: "capacitors"
    url: "https://www.smoltek.com/topic/capacitors.md"
  - name: "cnf-mim"
    url: "https://www.smoltek.com/topic/cnf-mim.md"
  - name: "semiconductor"
    url: "https://www.smoltek.com/topic/semiconductor.md"
  - name: "sustainability"
    url: "https://www.smoltek.com/topic/sustainability.md"
  - name: "sustainabletech"
    url: "https://www.smoltek.com/topic/sustainabletech.md"
---

# Smoltek Semi & AI

## [](https://www.smoltek.com#today-everyone-is-talking-about-ai-but-what-exactly-is-ai)Today, everyone is talking about AI. But what exactly is AI?

This arti­cle digs into the devel­op­ment of the “myth­i­cal” AI – which is on every­one’s lips today – from its begin­nings to the present day, and a lit­tle fur­ther into the near future. It also pro­vides an update on how it is pow­ered (the “AI chips”) – and what require­ments future capac­i­tors will need to meet to sup­port the ever-increas­ing pow­er loads that AI requires to function.

So, buck­le up. Here we go!

First, AI is not new – there have been ideas and hard­ware imple­men­ta­tions since the 1950s. But it is the com­bi­na­tion of deep learn­ing, big data and new chip archi­tec­tures that have explod­ed devel­op­ment over the past decade that has led to the heavy impact AI has had just in the last cou­ple of years. And the next 2–5 years will be about ener­gy effi­cien­cy, inte­gra­tion and spe­cial­iza­tion to run even larg­er mod­els – both in the cloud and at the edge.

## [](https://www.smoltek.com#the-ai-roadmap)The AI roadmap

### AI has been around in electronics/​microelectronics for quite some time.

- 1950s–1970s: The first ideas about AI were for­mu­lat­ed (Tur­ing, Rosenblatt’s per­cep­tron). But the hard­ware was too slow.
- 1980s–1990s: Wave of “expert sys­tems” and sim­pler ML algo­rithms. On the hard­ware side, CPUs and spe­cial­ized DSPs were used for sim­pler pat­tern recog­ni­tion, speech recog­ni­tion, and sig­nal processing.
- 2000s: GPUs began to be used for par­al­lel com­put­ing, giv­ing AI a major boost (espe­cial­ly in image recognition).
- 2010s: Deep learn­ing break­through (e.g. AlexNet 2012) – dri­ven by the com­bi­na­tion of large data sets + pow­er­ful GPUs + improved algo­rithms. This also saw the first ded­i­cat­ed AI chips (e.g. Google TPU 2016).

### AI development since the start

- From rules to learn­ing: Ear­ly AI sys­tems were rule-based (“if X then Y”). Mod­ern sys­tems learn them­selves from data.
- From small to gigan­tic mod­els: Net­works with a few thou­sand para­me­ters in the 90s → today hun­dreds of bil­lions of para­me­ters (e.g. GPT‑4, LLa­MA, PaLM).
- From niche to gen­er­al: AI is used today in smart­phones, cars, health­care, indus­try, lan­guage mod­els – everywhere.

### Drivers for today’s AI chip development

1. Expo­nen­tial­ly growth of data sets – AI mod­els need more data.
2. New algo­rithms (deep learn­ing, trans­form­ers) – more com­pu­ta­tion­al­ly intensive.
3. Lim­i­ta­tions in CPU/​GPU archi­tec­tures → need for spe­cial­ized hard­ware (TPU, NPU, FPGA, ASIC).
4. Moore’s Law slows down → instead, they focus on het­ero­ge­neous chips, chiplets, 3D inte­gra­tion and ener­gy efficiency.
5. Mar­ket dri­ving force – huge demand from com­pa­nies like Ope­nAI, Google, NVIDIA, Meta and others.

### What does the future (2–5 years ahead) look like for AI chips?

- Less ener­gy con­sump­tion per oper­a­tion: Ener­gy effi­cien­cy (TOPS/​Watt) becomes the biggest bottleneck.
- Het­ero­ge­neous sys­tems: Com­bi­na­tion of CPU + GPU + NPU/​TPU on the same package.
- Chiplets & advanced pack­ag­ing: 2.5D/3D inte­gra­tion, HBM (High Band­width Mem­o­ry) close to the com­pute cores.
- Spe­cial­iza­tion: Dif­fer­ent chips for dif­fer­ent tasks (train­ing vs infer­ence, data cen­ter vs edge).
- Mate­ri­als and new com­po­nents: Thin film capac­i­tors (like CNF-MIM), resis­tive RAM, pho­ton­ics for AI acceleration.
- Edge AI: Small, ener­gy-effi­cient AI chips in mobiles, IoT and vehi­cles – not just in giant data centers.

## [](https://www.smoltek.com#moores-law-and-the-development-of-ai-hardware)Moore’s Law and the development of AI hardware

![Transisto evolution 1950](https://www.smoltek.com/wp-content/uploads/2025/11/transisto-evolution-1950-2025-1200x800.webp)

*Gen­er­al overview of proces­sor devel­op­ment 1950–2025.*

### 1950–2000 (CPU era)

- Moore’s Law in full force: Num­ber of tran­sis­tors dou­bles every two years → expo­nen­tial improve­ment in com­put­ing power.
- AI research lim­it­ed to aca­d­e­m­ic exper­i­ments (sym­bol­ic AI, perceptrons).
- Hard­ware: CPUs (ser­i­al computing).

### 2000–2010 (GPU era)

- Moore’s Law con­tin­ues → but ener­gy con­sump­tion and clock fre­quen­cies reach limits.
- Par­al­lelism (GPUs) becomes the solu­tion for deep learn­ing → trains e.g. AlexNet (2012).
- The start of AI as prac­ti­cal­ly use­ful in image recog­ni­tion, lan­guage, games.

### 2010–present (TPU/​NPU era)

- Moore’s Law is start­ing to slow down: tran­sis­tor sizes <10 nm, fre­quen­cy increas­es are marginal.
- Solu­tion: spe­cial­iza­tion → TPU (Google), NPU (Apple Neur­al Engine), AI-ASICs (Graph­core, Cerebras).
- Focus on ener­gy effi­cien­cy per oper­a­tion (TOPS/​Watt) instead of just more transistors.
- AI mod­els are explod­ing in size (GPT, BERT, etc.).

### Future (next-gen AI chip, 2025–2030)

- Moore’s Law is on the verge to reach its phys­i­cal lim­its (sub 2 nm).
- Het­ero­ge­neous chips (CPU+GPU+NPU), chiplets and 3D integration.
- Advanced mem­o­ries (HBM, MRAM, RRAM).
- New mate­ri­als & com­po­nents (CNF-MIM, pho­ton­ics, neuromorphic).
- Focus on spe­cial­ized archi­tec­ture for dif­fer­ent AI work­loads: train­ing vs. infer­ence, data cen­ter vs. edge.

### From 1,000 transistors per chip in 1970 to over 50 billion transistors today.

![Moore's Law Transistor Count 1970](https://www.smoltek.com/wp-content/uploads/2025/11/moores-law-transistor-count-1970-2020.webp)

*50 years of CPU, tran­sis­tor count devel­op­ment. A log­a­rith­mic graph show­ing the time­line of how tran­sis­tor counts in microchips are almost dou­bling every two years from 1970 to 2020; also known as Moore’s law.*  
*Source: https://commons.wikimedia.org/wiki/File:Moore%27s\_Law\_Transistor\_Count\_1970-2020.png*

Moore’s Law states that the num­ber of tran­sis­tors will dou­ble every sec­ond year, roughly.This pre­dic­tion has cre­at­ed a glob­al need for high­ly advanced and high­ly minia­tur­ized capac­i­tors. Due to the enor­mous scal­ing of tran­sis­tors in advanced proces­sor chips, the need for decou­pling capac­i­tors mount­ed extreme­ly close to the proces­sor chip has explod­ed in recent years. 

How­ev­er, tra­di­tion­al capac­i­tor tech­nolo­gies have dif­fi­cul­ties deliv­er­ing high enough per­for­mance in a small enough form factor.

## [](https://www.smoltek.com#cnf-mim-to-the-rescue)CNF-MIM to the rescue

Smoltek’s patent pro­tect­ed nan­otech­nol­o­gy enables next-gen­er­a­tion capac­i­tors with extreme­ly high elec­tri­cal per­for­mance in a very small form factor.

Smoltek’s capac­i­tors, called CNF-MIM (Car­bon Nanofiber Met­al-Insu­la­tor-Met­al), can be placed direct­ly adja­cent to the cir­cuits they are to sup­port – the key to the rapid­ly grow­ing mar­ket for decou­pling capac­i­tors for high-end proces­sors – such as AI, High-Per­for­mance Com­put­ing and edge devices.

CNF-MIM capac­i­tors meets the semi­con­duc­tor indus­try’s demands for next-gen­er­a­tion capac­i­tors with unique characteristics: 

- Extreme­ly high capac­i­tance* per unit area
- Extreme­ly low elec­tri­cal losses
- Sev­er­al place­ment options clos­er to the proces­sor than com­pet­ing technologies
- Ener­gy effi­cien­cy (low ESR/​ESL, low leakage)
- Sta­bil­i­ty and life­time (break­down volt­age, reliability)

** Capac­i­tance is the abil­i­ty of a com­po­nent or cir­cuit to col­lect and store ener­gy in the form of an elec­tri­cal charge.*

![CNF MIM AI chip](https://www.smoltek.com/wp-content/uploads/2025/11/cnf-mim-ai-chip.webp)

*Mod­el of a “CNF-MIM AI chip”.*

## [](https://www.smoltek.com#the-four-key-parameters-for-ai-capacitors)The four key parameters for “AI capacitors”

Capac­i­tors intend­ed for use in next-gen­er­a­tion AI chips and high-per­for­mance com­put­ing tech­nol­o­gy need to be able to han­dle the following: 

### 1. Low ESR ([Equivalent Series Resistance](https://www.google.com/search?client=safari&rls=en&q=Equivalent%20Series%20Resistance&ie=UTF-8&oe=UTF-8&mstk=AUtExfAYnkZml5AgR4-IKcT_VaUqe3LZnqpwMSkgvr4Yc8YvQAv99h4M6WpTnzGbQ8pe1GVw1NZj8JvQrwBGrKNdeXckBFOWpXCpAxRHT-orjZ9EGEGGWNyce8wMQVWtrCAcsahxcgBVa3L3XUudew7B5_mtleInfEghxWLbx7AmKP-GxiM&csui=3&ved=2ahUKEwiElaWat6aRAxUvKBAIHYjjCwUQgK4QegQIARAD)) and ESL ([Equivalent Series Inductance](https://www.google.com/search?client=safari&rls=en&q=Equivalent%20Series%20Inductance&ie=UTF-8&oe=UTF-8&mstk=AUtExfAYnkZml5AgR4-IKcT_VaUqe3LZnqpwMSkgvr4Yc8YvQAv99h4M6WpTnzGbQ8pe1GVw1NZj8JvQrwBGrKNdeXckBFOWpXCpAxRHT-orjZ9EGEGGWNyce8wMQVWtrCAcsahxcgBVa3L3XUudew7B5_mtleInfEghxWLbx7AmKP-GxiM&csui=3&ved=2ahUKEwiElaWat6aRAxUvKBAIHYjjCwUQgK4QegQIARAE))

- ESR and ESL must be kept extreme­ly low for the capac­i­tor to deliv­er sta­ble and fast cur­rent dis­si­pa­tion at high frequencies.
- AI and HPC proces­sors oper­ate with bil­lions of tran­sis­tors that switch extreme­ly fast → they require clean and sta­ble sup­ply volt­age with­out interference.
- Low ESR/​ESL means fast response, less heat gen­er­a­tion and high­er ener­gy efficiency.

### 2. Breakdown voltage & reliability

- The capac­i­tor must with­stand the volt­age lev­els used in advanced process nodes (often around 0.7–1.2 V).
- An insuf­fi­cient­ly robust dielec­tric can lead to breakdown.
- Reli­a­bil­i­ty (dura­bil­i­ty) is cru­cial for data cen­ters and AI train­ing, where 24⁄7 oper­a­tion for years is the norm.

### 3. High capacitance density

- In today’s chips, sur­face area is extreme­ly valu­able – every square microm­e­ter must be uti­lized to its fullest.
- A capac­i­tor with high capac­i­tance den­si­ty makes it pos­si­ble to build more decou­pling capac­i­tors close to the proces­sor cores with­out tak­ing up too much space.
- This pro­vides local ener­gy stor­age that can deliv­er fast pow­er puls­es when mil­lions of cores are acti­vat­ed simul­ta­ne­ous­ly, which is cen­tral to AI mod­els that require mas­sive parallelism.

### 4. Relevance for AI and HPC

- AI chips (e.g. GPUs, TPUs) run matrix cal­cu­la­tions on a huge scale, with extreme­ly high and fast cur­rent variations.
- HPC chips must run at high fre­quen­cy with max­i­mum stability.
- In both cas­es, pow­er integri­ty (PI) becomes a bot­tle­neck – you have to be able to keep the volt­age sta­ble even though the cur­rent varies on a nanosec­ond scale.
- Here, the thin form fac­tor and elec­tri­cal prop­er­ties of the CNF-MIM capac­i­tor are par­tic­u­lar­ly attrac­tive: they can be placed close to the tran­sis­tors in the chip instead of on the cir­cuit board, which reduces par­a­sitics and pro­vides faster response.

## [](https://www.smoltek.com#the-three-capacitor-technologies-in-the-ai-race)The three capacitor technologies in the AI race

Now, it’s fair to say that CNF-MIM is not the only tech­nol­o­gy for decou­pling and ener­gy stor­age in next-gen­er­a­tion advanced chips. It’s com­pet­ing with two oth­er big capac­i­tor tech­nolo­gies; Trench Sil­i­con Capac­i­tors, or Deep Trench Capac­i­tors (TSC/​DTC) and Mul­ti­lay­er Ceram­ic Capac­i­tors (MLCC). The three tech­nolo­gies are all used for decou­pling and ener­gy stor­age, but they dif­fer great­ly in design, per­for­mance and integration. 

Let’s take a look.

![CNF MIM caps competition](https://www.smoltek.com/wp-content/uploads/2025/12/cnf-mim-caps-competition-1200x329.webp)

*CNF-MIM, TSC (or DTC) and MLCC. Illus­tra­tion of build­ing structure.*

### CNF-MIM (Carbon Nanofiber – Metal-Insulator-Metal)

- Struc­ture:
  
  - Based on ver­ti­cal car­bon nanofibers that pro­vide a 3D struc­ture and very large sur­face area in a small vol­ume. A sin­gle dielec­tric lay­er (one ALD step) is placed between the nanofibers and met­al contacts.
- Ben­e­fits:
  
  - Extreme­ly thin, can be inte­grat­ed direct­ly into the chip or on an interposer.
  - Very high capac­i­tance den­si­ty per area thanks to the nanostructure.
  - Low ESR/​ESL because the place­ment can be very close to the tran­sis­tors → fast tran­sient response.
- Chal­lenges:
  
  - Rel­a­tive­ly new tech­nol­o­gy needs to prove long-term reli­a­bil­i­ty and indus­tri­al­iza­tion on a large scale.

### TSC (Trench Silicon Capacitors) /​ DTC (Deep Trench Capacitors)

- Struc­ture:
  
  - Built by etch­ing “trench­es” (deep depres­sions) in sil­i­con and fill­ing them with dielec­tric and met­al in sev­er­al ALD steps. Uti­lizes sil­i­con as a substrate.
- Ben­e­fits:
  
  - Can be inte­grat­ed into the semi­con­duc­tor process or into sil­i­con interposer.
  - Pro­vides rel­a­tive­ly high capac­i­tance den­si­ty, espe­cial­ly com­pared to pla­nar MIM capacitors.
  - Good reli­a­bil­i­ty and proven technology.
- Chal­lenges:
  
  - Capac­i­tance den­si­ty is low­er than what CNF-MIM can achieve.
  - Process com­plex­i­ty (deep etch­ing) may lim­it cost-effectiveness.

### MLCC (Multilayer Ceramic Capacitors)

- Struc­ture:
  
  - Con­sists of mul­ti­ple lay­ers of ceram­ic dielectrics and met­al lay­ers stacked in pack­ages. Mount­ed as dis­crete com­po­nents on print­ed cir­cuit boards (PCBs).
- Ben­e­fits:
  
  - Cheap, mass-pro­duced and wide­ly avail­able technology.
  - High capac­i­tance in small vol­umes for dis­crete components.
- Dis­ad­van­tages:
  
  - Rel­a­tive­ly high ESR/​ESL due to place­ment fur­ther from the proces­sor (on the PCB, not in the chip).
  - Capac­i­tance varies strong­ly with tem­per­a­ture, volt­age and aging (espe­cial­ly with X7R/​X5R ceramics).
  - Dif­fi­cult to scale down to inte­gra­tion direct­ly into the chip.

### Technological differences in summary

|                           |                                  |                                            |                                            |
|---------------------------|----------------------------------|--------------------------------------------|--------------------------------------------|
| **Char­ac­ter­is­tic**    | **CNF-MIM**                      | **TSC /​ DTC**                             | **MLCC**                                   |
| **Place­ment**            | Direct­ly under chip/​interposer | In silicon/​interposer                     | On PCB                                     |
| **Capac­i­tance density** | Very high (nanos­truc­ture)      | High (trench structure)                    | Mod­er­ate (bulk)                          |
| **ESR/​ESL**              | Very low                         | Low                                        | Rel­a­tive­ly high                         |
| **Inte­gra­tion**         | Future tech­nol­o­gy for AI/​HPC | Already estab­lished in advanced packaging | Stan­dard com­po­nent, not chip-integrated |
| **Matu­ri­ty**            | In Devel­op­ment                 | Very mature and mass-produced              | Very mature and mass-produced              |

- MLCCs are cheap but too “slow” and too far away from the proces­sor. They have served the ear­ly days of AI but is now on the verge to be obsolete.
- TSC/​DTC is a proven solu­tion for inte­grat­ed capac­i­tance but lim­it­ed in den­si­ty. Serves today’s AI development.
- CNF-MIM capac­i­tors have the poten­tial to pro­vide the next leap in capac­i­tance den­si­ty and inte­gra­tion, mak­ing them par­tic­u­lar­ly attrac­tive for AI and HPC chips where every nanosec­ond and every square mil­lime­ter counts.

In con­clu­sion, by pio­neer­ing car­bon nan­otech­nol­o­gy Smoltek has a tru­ly dis­rup­tive capac­i­tor tech­nol­o­gy that enables the com­bi­na­tion of extreme­ly high elec­tri­cal per­for­mance in a high­ly minia­tur­ized prod­uct. Smoltek Semi­’s CNF-MIM capac­i­tor address­es key chal­lenges for next-gen­er­a­tion AI, high-per­for­mance com­put­ing and edge devices.

## [](https://www.smoltek.com#what-does-ai-have-to-say-about-this)What does AI have to say about this?

Okay, so this has been a deep dive into the devel­op­ment of AI and what is required of the capac­i­tors to keep this, not-so-new, thing rolling.

But the first ques­tion – what is AI wasn’t real­ly answered, was it? That might take some time to explain though. Might as well ask the AI about it.

### *“Hello, what is AI?”*

“Thanks for ask­ing. AI, or Arti­fi­cial Intel­li­gence, refers to com­put­er sys­tems designed to per­form tasks that typ­i­cal­ly require human intel­li­gence, such as learn­ing, prob­lem-solv­ing, deci­sion-mak­ing, and lan­guage under­stand­ing. It uses algo­rithms to ana­lyze data, iden­ti­fy pat­terns, and make deci­sions to achieve spe­cif­ic goals, with com­mon exam­ples includ­ing vir­tu­al assis­tants, rec­om­men­da­tion sys­tems, and self-dri­ving cars.”

### *“Alright, and what is HPC, or high-performance computing?”*

“High-per­for­mance com­put­ing (HPC) uses aggre­gat­ed com­put­ing pow­er, such as super­com­put­ers or clus­ters of com­put­ers, to solve large, com­plex prob­lems in sci­ence, engi­neer­ing, and busi­ness that would be impos­si­ble on a typ­i­cal com­put­er. HPC sys­tems process mas­sive datasets and per­form cal­cu­la­tions at extreme­ly high speeds, enabling com­plex mod­el­ing, data analy­sis, and arti­fi­cial intel­li­gence appli­ca­tions that dri­ve inno­va­tion and pro­vide sig­nif­i­cant com­pet­i­tive advantages.”

### *“And also, what are edge devices?”*

“Edge devices are hard­ware com­po­nents, like routers, sen­sors, smart­phones, and gate­ways, that con­nect to a core net­work and process data at or near its source, rather than send­ing it to a cen­tral­ized cloud or data cen­ter for analy­sis. These devices are essen­tial for edge com­put­ing, enabling real-time data pro­cess­ing, reduc­ing laten­cy, and improv­ing net­work effi­cien­cy in appli­ca­tions rang­ing from the Inter­net of Things to indus­tri­al automa­tion and smart cities.”

So, there you have it. An overview of AI devel­op­ments and how Smoltek’s CNF-MIM capac­i­tor tech­nol­o­gy has the poten­tial to become an impor­tant pil­lar of ener­gy-effi­cient pow­er man­age­ment in future chips.

* * *

Want to learn more about the extreme­ly small and ultra-thin CNF-MIM capac­i­tor? [Go here!](https://www.smoltek.com/smoltek-semi/capacitors/)