HBM Under Threat? Are Microsoft, Intel & Marvell Changing the AI Memory Game — Transcript
Full transcript
- 0:00Hey guys, welcome back to Market Signal.
- 0:02Hope you're doing well. It's another
- 0:03late night here in Phoenix, but there
- 0:05are a few key updates that I want to
- 0:07walk each and every one of you through
- 0:09to make sure that you stay informed and
- 0:11updated in this memory sector. You know,
- 0:13in my video last night, we talked about
- 0:15the Nvidia leadership shock and what it
- 0:18meant for the $500 billion AI buildout.
- 0:22But today on Saturday, November 14th, we
- 0:25are seeing a new batch of supply chain
- 0:28reporting that is dropping. And if you
- 0:30own Micron, SKH, and Samsung, you need
- 0:32to pay close attention to the bigger
- 0:35architectural picture that is emerging
- 0:37around us right now. There were three
- 0:39separate announcements over the last 48
- 0:41hours that are showing us some of the
- 0:43clearest signals yet that hyperscalers
- 0:46and chip designers are continuing to
- 0:49pursue a way to engineer around the HBM
- 0:52bottleneck that we see in the
- 0:53marketplace. For example, Microsoft just
- 0:56signed with Cerebrus. They dropped their
- 0:58earnings on the street at the end of
- 1:00this week. I'm going to show you that
- 1:01and what it means for the sector. And
- 1:03guys, I actually think Cerebrus might be
- 1:05at a really great entry point right now.
- 1:07But we're going to walk through all that
- 1:08data together in this video. And Intel
- 1:11is also pitching an HBM free inference
- 1:14GPU architecture. And then finally,
- 1:16Marll just laid out a blueprint showing
- 1:18that AI workloads are outgrowing the
- 1:21memory inside individual racks. They are
- 1:24now requiring optical memory fabrics and
- 1:27CXL expansion just to keep compute fed.
- 1:30So tonight, we're going to connect all
- 1:32of these dots and these signals that we
- 1:34see in the market. I'm going to separate
- 1:36the real architectural threats to HBM
- 1:39from the early stage hype and we're
- 1:41going to break down what it means to
- 1:43Micron and the other memory companies.
- 1:45So, make sure you hit like and make sure
- 1:47you subscribe to Market Signal. Guys, I
- 1:49hope you're having a great day. Let's
- 1:50dive in and let's talk about all of the
- 1:52signals and the architectural updates
- 1:55that we are that we need to know to stay
- 1:57informed in this sector overall. All
- 1:59right, guys. The first thing I want to
- 2:00jump to and talk about is the cerebrous
- 2:03earnings that they dropped on the
- 2:04street. So, let's walk through that
- 2:06together. I'm going to pull up their
- 2:07earnings statement here, and we'll walk
- 2:09through some of the key details. You can
- 2:11see the piece that I highlighted here.
- 2:13Surging demand for high-speed inference
- 2:15fuels growth from new and existing
- 2:18customers. So, they dropped their a
- 2:20record revenue in the third quarter of
- 2:222026. In addition, Cerebrus, they raised
- 2:25their fullear outlook overall. But a
- 2:29couple elements they their c as their
- 2:31CEO mentioned Andrew Feldman who's the
- 2:33co-founder and the CEO of Cerebrus he
- 2:35goes on record to say that Cerebrris
- 2:37technology remains the undisputed
- 2:40champion of speed. He went further on to
- 2:42say that they've exceeded our guidance
- 2:44across the the board for the third
- 2:46quarter for a third quarter in a row and
- 2:48delivered sequential improvements on all
- 2:51core financial metrics. And we we won
- 2:54new marquee customers including
- 2:56hyperscalers, neoclouds, and large
- 2:58enterprises. And we saw our pipeline for
- 3:00high-speed AI inference increase by more
- 3:02than 10x in the quarter. Take a look at
- 3:05a few of the metrics overall. So I'm
- 3:07going to show you their non-GAAP, their
- 3:09core financial uh results. the non-GAAP
- 3:11results. You can see in the first part I
- 3:13highlighted here the core cloud and
- 3:15other services revenue 140 almost 145
- 3:20million. It's up 272% year-over-year.
- 3:23But take a look at core total revenue.
- 3:25It's 231 million. It's up 70%
- 3:29year-over-year. And they have a core
- 3:30gross margin of 41%. And that's both
- 3:34sequ it's it's up both sequentially and
- 3:36year-over-year.
- 3:38But one of the other key things I want
- 3:40to show you is not only their remaining
- 3:42performance obligations, but what they
- 3:44said in the business highlights first.
- 3:47Two 26 billion in remaining performance
- 3:49obligations as of September 30th of this
- 3:52year. But really the piece that jumped
- 3:54out to me was this business highlights
- 3:56and the customer. So they went on record
- 3:58to say that Cerebrus is expand is
- 4:01rapidly expanding its customer base,
- 4:03broadening its reach and diversifying
- 4:05its revenue. So Oracle is integrating
- 4:07Cerebrra's fast inference into Oracle
- 4:09cloud infrastructure. But it's this
- 4:11second bullet point that really piqued
- 4:13my curiosity when the CEO said it in the
- 4:16in the analyst call. He said that
- 4:18Cerebrris announced an important cloud
- 4:20deal with a top three hyperscaler. More
- 4:23will be shared next year. It made me
- 4:25think who they're going to be working
- 4:27with because we already see Cerebrris
- 4:28working with Open AI, Amazon, Oracle.
- 4:32who would be the next top hyperscaler
- 4:34that they they would be working with as
- 4:36we move into 2027. And I saw that just
- 4:39the other day as it popped up. Excuse
- 4:41me, just saw it in another article as it
- 4:43popped up, but you can see it is
- 4:44Microsoft. So Cerebrus raises its 2026
- 4:48forecast and signs Microsoft as a cloud
- 4:51customer. We see cerebrus they expect up
- 4:53to 960 million in core revenue this year
- 4:57and its CEO says the Microsoft deal was
- 4:59signed actually in October. So if we go
- 5:02down into some of the details, the
- 5:04company now expects core revenue of 940
- 5:06to 960 960 million this year. And the
- 5:10results the results statement named a
- 5:13new customer only as a top three
- 5:15hyperscaler as I showed you which piqu
- 5:17my curiosity. But on an interview with
- 5:20CNBC, the chief executive Andrew
- 5:22Feldman, he mentioned that it is
- 5:24actually Microsoft which signed in
- 5:26October and he expects the revenue from
- 5:28the deal to start arriving in the first
- 5:30quarter of 2027. As I mentioned,
- 5:32Cerebras already works with Amazon with
- 5:35or Oracle and we know they work with
- 5:37Open AI which is a big part of that uh
- 5:40remaining performance obligation for the
- 5:42company. But we are seeing in the fourth
- 5:44quarter they forecast their revenue now
- 5:46to be 310 to 330 million. The analysts
- 5:51were expecting 271. So this is
- 5:54significant guidance above what the
- 5:56analysts expected for the upcoming
- 5:58quarter. Much of the growth rests with
- 6:00open AAI. So it's that compute deal that
- 6:03they have that underpins Cerebrus's 26
- 6:06billion in remaining performance
- 6:07obligations. the contract of revenue
- 6:10that is not yet booked according to
- 6:12Reuters.
- 6:13But there's a few other things I want to
- 6:15walk through, guys, with the deal. I
- 6:16actually think this is a very
- 6:18encouraging and positive earnings that
- 6:20Cerebras dropped on the street. And why
- 6:22does it matter so much that they won a
- 6:24new hyperscaler contract with Microsoft
- 6:28this cloud contract? Why does it matter?
- 6:30Why does it matter to Micron and SKH
- 6:32Heinix, the shareholders or investors in
- 6:34those companies? because we know how
- 6:37different the physics and the chip
- 6:38architectures are between these two
- 6:40companies. And just to remind you, I can
- 6:42pull it up because we know when you take
- 6:44a look at the difference and this
- 6:46graphic shows it perfectly. It's the
- 6:48memory bottleneck solve wafer scale
- 6:50versus the GPU mode. You're looking on
- 6:52the left side, you see the GPU plus the
- 6:54HBM cluster. So this is the Nvidia the
- 6:57AMD model where we have stacks of high
- 7:00bandwidth memory that are clustered
- 7:02right next to the compute. It is me
- 7:04according to the article here by uh by
- 7:07cerebrus it is memory traffic. It's
- 7:09memory bound and you could see the
- 7:11traffic and that has to data has to move
- 7:13back and forth from those stacks to the
- 7:15compute. Overall the cost of the bill of
- 7:17materials in this model is about 60 to
- 7:2070% for memory and we know that an HBM3
- 7:23stack cost about $300 each. If you look
- 7:25on the right side you can see Cerebrus'
- 7:28wafer scale engine that chip on a disk
- 7:30approach that they have. It's a very
- 7:32different architecture overall. They use
- 7:34this massive wafer scale engine with
- 7:37huge amounts of SRAMM integrated
- 7:39directly on chip. And by keeping memory
- 7:42directly on the silicon wafer, Cerebrus
- 7:45avoids the external HBM supply
- 7:47bottleneck and eliminates the energy and
- 7:49the latency penalty of moving data
- 7:52across off-chip interconnects for ultra
- 7:54fast AI inference.
- 7:56But we have to keep everything in
- 7:59perspective because as I just showed
- 8:00you, Cerebrus is at under 1 billion in
- 8:03annual revenue. It is still just a
- 8:05fraction of what Nvidia or Micron's
- 8:08footprint is. However, the microscope,
- 8:11excuse me, and this Microsoft deal, it
- 8:14doesn't mean that the hyperscalers are
- 8:16abandoning GPU clusters. However, what
- 8:19it does show us is it proves that when
- 8:21HBM supply is as tight as it is in the
- 8:24market and as expensive it is with the
- 8:26memory manufacturers having all the
- 8:28pricing power right now, the largest
- 8:31cloud buyers in the world are willing to
- 8:33fund alternative architectures to get
- 8:36inference capacity online. And I think
- 8:38this is what this Microsoft deal is
- 8:40showing us right now with with uh
- 8:42Cerebrus is they are working to try to
- 8:45get capacity online in many different
- 8:47ways. They're looking at alternative
- 8:49architectures. But as I said, Cerebrus's
- 8:52uh uh revenue is still under a billion.
- 8:54It it pales in consideration to the
- 8:56footprint that we see with Micron,
- 8:58Samsung, SKHEX, even Nvidia and AMD
- 9:01overall. But so I wanted to share that
- 9:03update with you because there's a couple
- 9:05things that I want to jump to next is
- 9:06just to take a look at uh Cerebras in
- 9:09general and is it a deal right now for
- 9:11us as investors? If I jump over to take
- 9:13a look at how the stock is trading, we
- 9:15know it had a couple good days to end
- 9:17the week. It was actually up by over 11%
- 9:19on Friday. It closed at a at 10872. But
- 9:22if you look to the right, you can see we
- 9:24are far off from where we are after the
- 9:27launch this year. When they had the IPO
- 9:29earlier this year, the stock got up to
- 9:31about $386.
- 9:33It reached 120 on Friday as the intraday
- 9:36high. But I want to call your attention
- 9:38to just what the fair price and
- 9:39ultimately what the analyst targets are.
- 9:41Now, fair price. Cerebrus is right in
- 9:44line with what the fair price is.
- 9:46According to the model on investing.com,
- 9:48you can see it's $106 a share. The stock
- 9:51is trading at about $109 a share right
- 9:53now. But analyst estimates have this
- 9:55stock at about $219 a share. $101%
- 10:00upside today from where we are trading.
- 10:02And just take a look at how Cerebras
- 10:04actually has been penalized compared to
- 10:06some of to the broader chip market from
- 10:08the from July to where we stand today.
- 10:11Cerebrus is in the orange line right
- 10:13here and it's comparing against the S&P
- 10:15500 as well as the chip index and you
- 10:17can see it's significantly down. This
- 10:19stock has been put under pressure way
- 10:23way further and harder than we see with
- 10:25the broader indexes overall. Was that
- 10:27pressure justified? We know it jumped
- 10:29out of the gate in its IPO. I stayed
- 10:31away from it in its IPO and and in full
- 10:33transparency, I'm not a shareholder in
- 10:35Cerebras at this moment. I've been
- 10:37waiting for the right price point to
- 10:40move in on the stock and I'm going to be
- 10:41honest with you, I think right now is a
- 10:43right price point for me. Make sure you
- 10:45do your own due diligence. Remember, I'm
- 10:47not a financial advisor. Make sure you
- 10:49do your own research and it fits your
- 10:51own individual risk tolerance, but where
- 10:53the stock is trading right now, the
- 10:55earnings that they dropped on the
- 10:56street, the deals that they continue to
- 10:58sign, the most recent deal they just
- 10:59signed with Microsoft, I do think we are
- 11:01in a position to consider this stock as
- 11:04an entry point at 108. And I just wanted
- 11:06to show you where the stock is sitting
- 11:08from the analyst perspective. So we have
- 11:1111 analysts that are following the stock
- 11:13all listed as a buy. As I mentioned, the
- 11:15average 12-month target is about 219.
- 11:18But you can see the range. We have
- 11:20everything from 170 a share for Rosenlat
- 11:23all the way up to 205 for Craig Harlem,
- 11:26which is representing anywhere between a
- 11:2856 to 130% premium for where the stock
- 11:31trades in the market today. So guys, I
- 11:33wanted to cover that first and foremost
- 11:35because we are seeing great earnings
- 11:36with Cerebras. We are seeing
- 11:38hyperscalers sign with them as well.
- 11:40It's just another strong sign for the
- 11:41demand, the memory demand that we have
- 11:43in the market. And the fact that we're
- 11:45seeing hyperscalers and other customers
- 11:47not only work with Micron, but trying to
- 11:49solve the the bottleneck by pursuing
- 11:51other architectures that we see in the
- 11:54market. But it wasn't just Cerebrus
- 11:56today too. We had other signals that I
- 11:58want toh show you around this
- 12:00architecture focus around what customers
- 12:02are doing to try to solve the high
- 12:04bandwidth bottleneck that we see in the
- 12:06market. There was one from Intel that I
- 12:08want to cover with you and then the last
- 12:10one from from Marll and then I want to
- 12:12show you a little bit of the supply side
- 12:14that we're going to see as we move into
- 12:162027. The fact that we are going to be
- 12:18continue to be constrained as we move
- 12:21into 27. But first, let's talk about
- 12:23what Digi Time showed with regards to
- 12:25Intel and what their approach is or what
- 12:27they're putting together right now with
- 12:29a partnership over in Korea. So, you can
- 12:31see Intel formed a 23 company Korean AI
- 12:35alliance as it pitches HBM free
- 12:38inference GPUs. So, Intel is courting
- 12:41South Korea's AI ecosystem on two
- 12:43fronts. The first is a 23 company
- 12:46alliance built around Korean-made
- 12:48servers and an inference GPU that swaps
- 12:51costly HBM for low power DRAM.
- 12:54That's the one kind of the front that I
- 12:57want to focus on from this article
- 12:59overall because what we are seeing is
- 13:01they announced the launch of a KAI full
- 13:04stack alliance describing it as a
- 13:06commitment to creating innovation
- 13:08together with Korean companies on a open
- 13:11infrastructure and ecosystem built
- 13:13around domestically manufactured
- 13:14servers. Inference an inference GPU
- 13:17without HBM. The centerpiece project was
- 13:21was not a CPU but Crescent Island. the
- 13:24inference GPU that Intel renamed the GPU
- 13:27D series in October. It uses LP DDR5X
- 13:32instead of HBM and draws 350 watts,
- 13:35slightly over half the 600 watts of
- 13:38Nvidia's RTX Pro 6000. Intel said memory
- 13:42capacity could reach 480 gigabytes, up
- 13:46to five times that an Nvidia's card. So
- 13:49that is a key piece that we are seeing
- 13:50with Intel, this alliance that they're
- 13:53they're launching right there in Korea.
- 13:54And I think for for us Micron investors
- 13:57or advanced memory investors, this is a
- 14:00particularly for Micron investors, this
- 14:01is a classic two-sided coin. On one
- 14:04hand, if HBM free inference chips gain
- 14:07traction for lighter enterprise
- 14:09workloads, it will reduce HBM content
- 14:12per accelerator. But on the other hand,
- 14:15deploying lower cost inference hardware
- 14:18at scale still consumes massive volumes
- 14:21of standard DDR and LP DDR bit capacity.
- 14:25So right now, this is just a proposed
- 14:27roadmap that we're seeing Intel try to
- 14:30put together with the Korean companies.
- 14:32It is not a proven commercial threat
- 14:35matching the HBM accelerator
- 14:37performance. So, I would treat it as a
- 14:39long-term watch list rather than an
- 14:41immediate order cut or an immediate
- 14:43threat that we see in the marketplace.
- 14:46And in addition to this, we're also
- 14:47seeing Marll come out. We know that
- 14:49Marll is going to be at at the at an
- 14:52upcoming conference. They're actually
- 14:53going to be at the super, it's called a
- 14:56supercomputing conference, the
- 14:58international conference for high
- 14:59performance computing. It's actually
- 15:01starting tomorrow. It runs from November
- 15:0315th through the 20th in Chicago. And
- 15:06what Marll came out today is as you can
- 15:09see I've have it highlighted here. They
- 15:10are mentioning that as AI workloads
- 15:12outgrow single rack memory and
- 15:14bandwidth. Marll demonstrates
- 15:16interconnect and memory technologies
- 15:18engineered to scale together. So as a
- 15:21leader in the connectivity and the
- 15:23memory innovation, Marll is addressing
- 15:25the connectivity challenges with
- 15:27advances in switch and interconnect
- 15:29technologies the and memory technologies
- 15:32with innovations in CXL memory expansion
- 15:35and acceleration. Additionally, Marll is
- 15:37pioneering technologies to unify these
- 15:40domains and scale memory and
- 15:42interconnect as an as a single system.
- 15:45So Marvel will showcase the technologies
- 15:48enabling the next generation of AI
- 15:50infrastructure across scale up, scale
- 15:53out and scale across architectures.
- 15:56Optical interconnect technologies may
- 15:58allow memory and compute to scale
- 16:00together across multiple racks, growing
- 16:03the size of the scale up domain. And the
- 16:05demonstrations that Marll will have at
- 16:08their booth, include the Marll photonic
- 16:10fabric memory module, the CXL memory
- 16:13solutions, including the Maravell
- 16:15structer, the addin card, the near
- 16:18memory compute and expansion and
- 16:19compression solutions, and partner
- 16:21memory demonstrations with Structura X
- 16:24memory expansion controllers featuring
- 16:26both DDR4 and DDR5. So what Marll is
- 16:31telling the market is really simple
- 16:33here. What they are saying is that the
- 16:34next phase of the AI buildout isn't just
- 16:37about stacking more HBM directly onto a
- 16:40single GPU. It's about disagregating
- 16:44memory and pooling it across racks using
- 16:46high-speed optical interconnects and CXL
- 16:49controllers. It expands the total
- 16:52addressable market for optical and
- 16:55networking players like Marll, Broadcom,
- 16:58Arista, Coherent, Lummentum, Credo. And
- 17:02it also proves that the HBM the excuse
- 17:04me the memory bandwidth remains the
- 17:06central bottleneck of AI scaling. So
- 17:09while all this is happening right now we
- 17:11are also seeing signs on the supply
- 17:14side. So we're the supply side is
- 17:16showing us that the HBM super cycle is
- 17:19heating up as we head into 2027 and
- 17:222028. Just take a look at a few of these
- 17:24pieces that I put together to just give
- 17:26you a demonstration of how this is
- 17:28heating up and what we are seeing from
- 17:29some key providers right now. You can
- 17:31see on your screen HBM supply heats up.
- 17:34The HBM super cycle's next phase for
- 17:372027 and 2028. On the left side, we know
- 17:40that SKHX is reportedly just pushed its
- 17:44HBM4 production yields past 70%. Giving
- 17:47it giving its stronger unit cost uh
- 17:50economics and volume output as HBM4
- 17:54ramps. In the middle, we're talking
- 17:56about Samsung and how it's accelerating
- 17:58its new HBM packaging facility in Asia
- 18:00by a full year. They are targeting June
- 18:032028, opening with an initial 4.48
- 18:07billion facility investment and roughly
- 18:1030 billion in subsequent advanced
- 18:12equipment spending discussed. And then
- 18:14finally on the right side we see CXMT in
- 18:17China is pushing into the higher value
- 18:20DDR5 server modules and the LP DDR DDR6
- 18:25products. And while CXMT still faces
- 18:27steep hurdles in commercial volume and
- 18:30US export constraints, it reinforces
- 18:33that conventional DRAM pricing won't
- 18:35stay immune from com from competition
- 18:38forever. And when you combine these
- 18:40three, you can see at the bottom the big
- 18:42picture. HBM supply is expanding with
- 18:44more capacity, higher yields, and new
- 18:48competition, setting up a more
- 18:49competitive memory market in 2027 and
- 18:522028. We, as I mentioned, we see
- 18:55Samsung's accelerated 30 billion plus
- 18:57packaging buildout for 2028. SKHik 70
- 19:02plus HBM4 yields. And the core debate
- 19:05overall for Micron, for us investors, it
- 19:09needs to shift in our mind. It is no
- 19:11longer about can we sell every wafer
- 19:13that the company has today. It is
- 19:15shifting. And what that shift is is
- 19:17continuing to show us is just how tight
- 19:20HBM pricing can stay once competing
- 19:23capacity hits its full stride in 2028.
- 19:26That's the central question. When we hit
- 19:28our full stride in 2028, can Micron and
- 19:31the other companies keep that pricing
- 19:33leverage? Will they be able to keep that
- 19:35HBM pricing, that premium pricing in the
- 19:37market? because we're going to see more
- 19:40capacity hit in the marketplace. But as
- 19:42of right now, we are still constrained
- 19:45as ever. We are going to be constrained
- 19:46in this market as we move into 2027 and
- 19:502028 despite what we are seeing from not
- 19:53only SKH, from CXMT, from Samsung, from
- 19:56Micron as well. Even what Cerebrus is
- 19:58doing right now. You think about how the
- 20:00other players are trying to work around
- 20:02the HBM bottleneck. I gave you two
- 20:04examples tonight. Intel as well as what
- 20:07Marll is talking about as well. So to
- 20:09put it all together in a clear kind of
- 20:11framework for us as we head into
- 20:13Monday's opening bell, first and
- 20:15foremost, we have to pay attention to
- 20:18Tuesday's catalyst moment with Nvidia
- 20:21when they step to the podium with their
- 20:22new leadership team. They share their
- 20:24earnings and they discuss what their
- 20:27future is going to be without Jensen.
- 20:28It's going to be the biggest moment next
- 20:30week. We have Enthropics IPO that will
- 20:32mean a ton for this sector as well. But
- 20:35we have to realize that HBM demand is
- 20:37not collapsing. It is just diversifying.
- 20:40Near-term HBM3 and HBM4 supply remains
- 20:43tightly sold out, which is why Micron's
- 20:46fundamentals remain strong despite the
- 20:48pullback we have seen in the sector
- 20:50recently. And secondly, watch the
- 20:53architectural hedges that we're seeing.
- 20:55Microsoft's deal with Cerebrus and
- 20:57Marll's CXL and optical memory push.
- 21:00They prove that hyperscalers are
- 21:02actively building complimentary ways to
- 21:04scale memory bandwidth outside of
- 21:07standard GPU plus HBM architectures. But
- 21:10guys, let me know in the comments below
- 21:12what you think, what your thoughts are.
- 21:14What did you think of Cerebrus's
- 21:16earnings? Do you think the wafer scale
- 21:18chip that Cerebras has in the market or
- 21:20maybe even the CXL memory pooling that
- 21:23Mel Marll is talking about, do you
- 21:25actually think that could dent the HBM
- 21:28growth rate that we may see in 2028 or
- 21:31will HBM continue to be the undisputed
- 21:33king in the AI compute? Guys, that's the
- 21:36update I have for tonight. I hope you
- 21:38enjoyed the video. Make sure you hit
- 21:39like. Make sure you join Market Signal
- 21:41if you haven't done it already. and I'll
- 21:43see you on the next
About this transcript
This page contains the full transcript of HBM Under Threat? Are Microsoft, Intel & Marvell Changing the AI Memory Game by Market Signal, generated from the public captions YouTube serves with the video. The transcript has 3,745 words across 553 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.