What the hell odes “idle normalized platform power” mean…?
It means they measured board power draw from the battery sensor, and deducted how much the device uses at idle.
This is basically gold standard testing that gets used in phones but not in laptops. Most laptop reviewers just blindly trust inaccurate package power figures.
Qualcomm massive hardware superiority is both real and expected. Qualcomm is competitive with Apple on the phone side. There is nothing x86 that is competitive with M series. So it’s self evident that a qualcomm SoC with their modern IP is not something x86 can handle.
You just have a very old (IP Wise) qualcomm chip. The reason you’re getting 10fps is that your iGPU is worse than a Qualcomm S25 Ultra iGPU while it’s thermally throttling.
Look up 8 Gen 2, that + OC is what you have in your laptop for iGPU. The 8 Gen 2 iGPU is a traditional phone iGPU designed for simple shaders and that struggles with deferred rendering. Adreno 830 is much more compatible with how games on windows render.
Qualcomm back is to the wall, Just like Intel. But X2 Elite and 8 Elite 5 show that Qualcomm are capable of great engineering that can rival Apple silicon.
If they keep it up, they will succeed. If they don’t, they will crumble. Live or die.
The first iteration of the Snapdragon Elite SoCs for Laptops had three major problems: 1. Qualcomm's overreliance on Microsoft to get the x86 emulator working as good as Apple's solution when Apple moved to their own CPUs (M series). Apple's emulator worked (IMHO) almost amazingly well, while the one for Qualcomm/Windows-on-ARM didn't. Made it easy for users to switch to Apple's own CPU designs.
2. Qualcomm relied way too much on their Adreno GPU architecture being the fastest in Smartphones; that simply didn't translate at all to Windows-on-ARM.
3. Qualcomm's biggest problem was that they were so fixated on "our SoCs are faster than Apple's M SoCs". Chasing (and missing ) that target, they ignored IMHO at their peril the niche that could have been theirs: highly efficient, light subnotebooks and 2-in-1s with great battery life. In that class, the SoC has to be just fast enough for a good user experience in routine office tasks, but especially allow for great battery life. Sort-of what the MacBook Air is for Apple's lineup.
Failing to do that when they launched the Elites, Qualcomm allowed Intel to come back with Lunar Lake (and AMD with Ryzen AI/Zen5) and for x86 to successfully defend that segment.
>> Qualcomm relied way too much on their Adreno GPU architecture being the fastest in Smartphones;
they didn’t.
They reused their 2022 iGPU (Adreno 740) in X Elite and OCed. The smartphone iGPU that they sold while selling X Elite (Adreno 830) was faster within a phone power envelope.
>> Focused too much on Apple
X Elite was far worse than Apple. X Elite was competitive with x86, and thus by definition it was not competitive with Apple in the slightest. Qualcomm IP only caught up to Apple in 2024 for GPU and 2025 for CPU. 2023 CPU + 2022 GPU in X Elite was not a recipe for success.
The Skyworks/Qorvo merger creates a fascinating dynamic for Apple's modem transition. QRVO's RF front end capabilities combined with SWKS antenna tunig expertise could actually help Apple accelerate their in-house modem timline. The $500M+ synergies are real but the bigger story is how this consolidation leaves Broadcom as the only other major RF player. I think the China aproval risk is underpriced here though, especially given how dependent both companies are on iPhone content. That ~$17 per iPhone RF content gives them incredible leverage but also makes them vulnerable to Apple's vertical integration push.
The failed Tower Semiconductor aquisition by Intel really was a pivotal moment for the US semiconductor landscape. Tower brings decades of expereince with specialty process technologies and external customer support that Intel desperately needs. Its particularly painful when you consider how Tower's analog and power management expertise could have accelerated Intel's foundry credibility. The geopolitical dynamics that killed this deal are complex, but the strategic loss is undeniable. Makes you wonder what alternative paths remain for Intel to build that external PDK capability organicly.
The "relatively safe" AI trio framing makes sense, but I think the real insight is how TSMC's position differs from Nvidia and Broadcom. Those two are design houses that need TSMC to execute - TSMC is the physical chokepoint.
What I find intresting is how much pricing power TSMC actually has vs. how much they're willing to exercise. They've been raising wafer prices gradually (3-5% annually in recent quarters), but they could probaly push harder given CoWoS demand and the 3nm/2nm ramp. The fact that they're not squeezing customers harder suggests they're playing a long game - keeping barriers to entry high for Samsung/Intel while maintaining custmer loyalty.
Your Intel 18AP section is fascinating. The 70% probability number feels about right based on what I've heard through the grapevine. The Tower Semi acquisition failure really was a critical missed opportunity for external PDK expertise. Even if 18AP works technically, the trust deficit with potential customers is massive after the 18A delays and issues.
One other thought: TSMC's overseas fab strategy (Arizona, Japan, Europe) is expensive but probably necessary insurance against geopolitical risks. Those fabs will never have the same cost structure as Taiwan, but they serve as a hedge that justifies the strategic investment even if ROIC is lower.
Really enjoyed the deep dive on the AI trio - the "safe" framing makes sense for anyone with a 2+ year horizon. Your point about TSMC being in the relatively safe category resonates, especialy given how they're basically the only game in town for leading-edge logic. What I find intresting about your Intel section (which I know is controversial!) is the PDK quality argument - you're absolutely right that even advanced process nodes can regress if the PDK is trash. The area vs performance tradeoffs you described when device models are missing or poorly characterized is spot on from my experince. Also your valuation approach using GloFo as a floor is creative - Intel at 1.2x book does seem like a reasonable floor given Trump's implicit backstop. The fiduciary duty argument to chip designers is fascinating - TSMC pricing power with zero competition is genuinely scary long-term. Great write-up!
The 'safe AI trio' framework is spot on - TSMC's positioning is incredibly defensable given the PDK ecosystem gap you detailed in the Intel section. What really resonates is your point about TSMC not just being about the silicon process itself but the entire design ecosystem that makes it work. The contrast between Intel's 18A PDK quality issues versus TSMC's mature library support perfectly ilustrates why technical specs alone don't determine winners in foundry. One thing I'd add to your TSMC bull case is the exponential token growth dynamics CC Wei keeps emphasizing, because even if HPC revenue stayed flat QoQ (which puzzled many), the utilization and pricing power implications for 2H26 could be huge. Also appreciate the disclosure section transparency and the patriotic Intel pitch even though you aknowledge all the biases - that kind of honest discussion about fiduciary duty to prevent TSMC monopoly pricing is something more board rooms should be having.
Really strong analysis across the board! Your point about the EDA majors being solid long-term holds but with rich valuations is exactly right. Cadence is in that rare position where it's simultaneously defensive (duopoly pricing power, mission-critical software) and offensive (AI design tools creating genuine incremental TAM). The chalenge is timing the entry - these stocks rarely get cheap, and waiting for a pullback might mean missing years of compounding. I'd argue that for non-leveraged accounts, both CDNS and SNPS are worth holding through volatility given the secular tailwinds. The switch from FinFET to GAA transistors alone guarantees elevated EDA spending for the next 3-5 years. Great read!
Didnt Intel sell the future revenue from some of their newer fabs to PE to reduce debt ? Not sure of the terms but even if the fab a great success it may not generate much cash for Intel ?
Just to be pedantic and repeat something I pointed out earlier, SMH is not quite market cap weighted.
First, any sector or industry fund cannot hold big winner names in the real proportion of their market cap due to SEC RIC capping rules. E.g. Compare the NVDA/INTC market cap ratio to their percentages in SMH ratio.
Second, SMH also uses trading volume in the ranking and weighting process. So it’s not just market cap.
The first explains why some non sector specific market wide funds can take big bets on winners and do well as the overweighting is easier to do without hitting the capping limits.
Great post. I think they can delay Intel Foundry's death, but it’s still inevitable. TSMC will eventually become a super-monopoly, and we’ll have to live with it.
What the hell odes “idle normalized platform power” mean…?
It means they measured board power draw from the battery sensor, and deducted how much the device uses at idle.
This is basically gold standard testing that gets used in phones but not in laptops. Most laptop reviewers just blindly trust inaccurate package power figures.
Qualcomm massive hardware superiority is both real and expected. Qualcomm is competitive with Apple on the phone side. There is nothing x86 that is competitive with M series. So it’s self evident that a qualcomm SoC with their modern IP is not something x86 can handle.
You just have a very old (IP Wise) qualcomm chip. The reason you’re getting 10fps is that your iGPU is worse than a Qualcomm S25 Ultra iGPU while it’s thermally throttling.
Look up 8 Gen 2, that + OC is what you have in your laptop for iGPU. The 8 Gen 2 iGPU is a traditional phone iGPU designed for simple shaders and that struggles with deferred rendering. Adreno 830 is much more compatible with how games on windows render.
Qualcomm back is to the wall, Just like Intel. But X2 Elite and 8 Elite 5 show that Qualcomm are capable of great engineering that can rival Apple silicon.
If they keep it up, they will succeed. If they don’t, they will crumble. Live or die.
The first iteration of the Snapdragon Elite SoCs for Laptops had three major problems: 1. Qualcomm's overreliance on Microsoft to get the x86 emulator working as good as Apple's solution when Apple moved to their own CPUs (M series). Apple's emulator worked (IMHO) almost amazingly well, while the one for Qualcomm/Windows-on-ARM didn't. Made it easy for users to switch to Apple's own CPU designs.
2. Qualcomm relied way too much on their Adreno GPU architecture being the fastest in Smartphones; that simply didn't translate at all to Windows-on-ARM.
3. Qualcomm's biggest problem was that they were so fixated on "our SoCs are faster than Apple's M SoCs". Chasing (and missing ) that target, they ignored IMHO at their peril the niche that could have been theirs: highly efficient, light subnotebooks and 2-in-1s with great battery life. In that class, the SoC has to be just fast enough for a good user experience in routine office tasks, but especially allow for great battery life. Sort-of what the MacBook Air is for Apple's lineup.
Failing to do that when they launched the Elites, Qualcomm allowed Intel to come back with Lunar Lake (and AMD with Ryzen AI/Zen5) and for x86 to successfully defend that segment.
You are ignoring what I said.
>> Qualcomm relied way too much on their Adreno GPU architecture being the fastest in Smartphones;
they didn’t.
They reused their 2022 iGPU (Adreno 740) in X Elite and OCed. The smartphone iGPU that they sold while selling X Elite (Adreno 830) was faster within a phone power envelope.
>> Focused too much on Apple
X Elite was far worse than Apple. X Elite was competitive with x86, and thus by definition it was not competitive with Apple in the slightest. Qualcomm IP only caught up to Apple in 2024 for GPU and 2025 for CPU. 2023 CPU + 2022 GPU in X Elite was not a recipe for success.
at this point samsung foundry seems much attractive than intel
The Skyworks/Qorvo merger creates a fascinating dynamic for Apple's modem transition. QRVO's RF front end capabilities combined with SWKS antenna tunig expertise could actually help Apple accelerate their in-house modem timline. The $500M+ synergies are real but the bigger story is how this consolidation leaves Broadcom as the only other major RF player. I think the China aproval risk is underpriced here though, especially given how dependent both companies are on iPhone content. That ~$17 per iPhone RF content gives them incredible leverage but also makes them vulnerable to Apple's vertical integration push.
The failed Tower Semiconductor aquisition by Intel really was a pivotal moment for the US semiconductor landscape. Tower brings decades of expereince with specialty process technologies and external customer support that Intel desperately needs. Its particularly painful when you consider how Tower's analog and power management expertise could have accelerated Intel's foundry credibility. The geopolitical dynamics that killed this deal are complex, but the strategic loss is undeniable. Makes you wonder what alternative paths remain for Intel to build that external PDK capability organicly.
The "relatively safe" AI trio framing makes sense, but I think the real insight is how TSMC's position differs from Nvidia and Broadcom. Those two are design houses that need TSMC to execute - TSMC is the physical chokepoint.
What I find intresting is how much pricing power TSMC actually has vs. how much they're willing to exercise. They've been raising wafer prices gradually (3-5% annually in recent quarters), but they could probaly push harder given CoWoS demand and the 3nm/2nm ramp. The fact that they're not squeezing customers harder suggests they're playing a long game - keeping barriers to entry high for Samsung/Intel while maintaining custmer loyalty.
Your Intel 18AP section is fascinating. The 70% probability number feels about right based on what I've heard through the grapevine. The Tower Semi acquisition failure really was a critical missed opportunity for external PDK expertise. Even if 18AP works technically, the trust deficit with potential customers is massive after the 18A delays and issues.
One other thought: TSMC's overseas fab strategy (Arizona, Japan, Europe) is expensive but probably necessary insurance against geopolitical risks. Those fabs will never have the same cost structure as Taiwan, but they serve as a hedge that justifies the strategic investment even if ROIC is lower.
Really enjoyed the deep dive on the AI trio - the "safe" framing makes sense for anyone with a 2+ year horizon. Your point about TSMC being in the relatively safe category resonates, especialy given how they're basically the only game in town for leading-edge logic. What I find intresting about your Intel section (which I know is controversial!) is the PDK quality argument - you're absolutely right that even advanced process nodes can regress if the PDK is trash. The area vs performance tradeoffs you described when device models are missing or poorly characterized is spot on from my experince. Also your valuation approach using GloFo as a floor is creative - Intel at 1.2x book does seem like a reasonable floor given Trump's implicit backstop. The fiduciary duty argument to chip designers is fascinating - TSMC pricing power with zero competition is genuinely scary long-term. Great write-up!
The 'safe AI trio' framework is spot on - TSMC's positioning is incredibly defensable given the PDK ecosystem gap you detailed in the Intel section. What really resonates is your point about TSMC not just being about the silicon process itself but the entire design ecosystem that makes it work. The contrast between Intel's 18A PDK quality issues versus TSMC's mature library support perfectly ilustrates why technical specs alone don't determine winners in foundry. One thing I'd add to your TSMC bull case is the exponential token growth dynamics CC Wei keeps emphasizing, because even if HPC revenue stayed flat QoQ (which puzzled many), the utilization and pricing power implications for 2H26 could be huge. Also appreciate the disclosure section transparency and the patriotic Intel pitch even though you aknowledge all the biases - that kind of honest discussion about fiduciary duty to prevent TSMC monopoly pricing is something more board rooms should be having.
Really strong analysis across the board! Your point about the EDA majors being solid long-term holds but with rich valuations is exactly right. Cadence is in that rare position where it's simultaneously defensive (duopoly pricing power, mission-critical software) and offensive (AI design tools creating genuine incremental TAM). The chalenge is timing the entry - these stocks rarely get cheap, and waiting for a pullback might mean missing years of compounding. I'd argue that for non-leveraged accounts, both CDNS and SNPS are worth holding through volatility given the secular tailwinds. The switch from FinFET to GAA transistors alone guarantees elevated EDA spending for the next 3-5 years. Great read!
The guy just hates qcom for no reason no cpu on market beats x elite 2
Didnt Intel sell the future revenue from some of their newer fabs to PE to reduce debt ? Not sure of the terms but even if the fab a great success it may not generate much cash for Intel ?
Are you still shorting Astera Labs ($ALAB)?
TSM and NVDA have risks of crashing along with AI bubble burst (who knows when). TSM also has China invasion risk too.
Thanks for the detailed write up. For a semi tourist this was especially helpful.
Just to be pedantic and repeat something I pointed out earlier, SMH is not quite market cap weighted.
First, any sector or industry fund cannot hold big winner names in the real proportion of their market cap due to SEC RIC capping rules. E.g. Compare the NVDA/INTC market cap ratio to their percentages in SMH ratio.
Second, SMH also uses trading volume in the ranking and weighting process. So it’s not just market cap.
The first explains why some non sector specific market wide funds can take big bets on winners and do well as the overweighting is easier to do without hitting the capping limits.
what do you see in TTMI and LITE?
Great post. I think they can delay Intel Foundry's death, but it’s still inevitable. TSMC will eventually become a super-monopoly, and we’ll have to live with it.