If there were accounting irregularities at $NVDA, then perhaps whoever was the CFO should resign now :)
Mo
Verified account This account is verified. Learn more@PersianMacroGuy
Living the island life as a trader on the beaches of Puerto Rico.
ex-software engineer: Google, Meta
ex-software engineer: Google, Meta
1K Following 6.2K Followers 2.5K Posts
Latest posts 5
buying $NVDA here for a trade
People don’t understand how much momentum Google has right now. This is not an accident. If I were a frontier lab, I would be panicking right now
TPUs are helping Google $GOOG win the AI race by enabling a model-silicon recursive self-improvement (RSI) loop.View quoted post on X
Up until recently, RSI was mostly a concept that applied to AI models helping make better AI models. I believe Google is changing that narrative. It's becoming clear that model-silicon co-design is the future of AI, and RSI has a lot more leverage in the model-chip feedback loop.
Because of scaling laws, the AI race has always been more of a compute acquisition race. Model design is important, but secondary to compute. More compute gets you better results, while better models help you extract more intelligence at lower costs. So the race really comes down to first who can accumulate compute at the cheapest price, and then who can use it most efficiently.
TPUs are arguably the biggest leverage Google has in the AI race. Internally, TPUs are 30-40%+ cheaper than comparable $NVIDIA setups, even after paying Broadcom and TSMC, since Google isn't paying NVIDIA's ~75% gross margins. When frontier AI companies are pouring hundreds of billions into compute, that gap means Google gets way more compute per dollar with much better ROIC, and that compounds every cycle.
Selling TPUs to outside buyers like Anthropic $ANT, Meta, and Oracle also lets Google spread its chip R&D across more volume. Because they can afford thinner margins than NVIDIA, they can undercut on price while still keeping the best economics for themselves.
But I'd argue those substantial cost savings aren't even the biggest edge for Google. As I said in my last post, I believe every frontier lab will eventually become a chip company because otherwise you leave too much efficiency on the table.
Because Google knows Gemini's roadmap in advance, they can strip out the extra features a GPU needs and design chips strictly around Gemini's bottlenecks. For example, TPU 8i triples on-chip memory and redesigns chip-to-chip networking specifically for long-context MoE reasoning, giving them up to 80% better performance per dollar than Ironwood.
On the flip side, the Gemini team designs the next model around the TPU's exact layout before the chip is even made. TPUs are much more rigid than GPUs, so outside models designed for NVIDIA can lose half their utilization in key layers on a TPU, while Gemini is tailored to fit it.
Outside buyers like Anthropic and Meta still buy TPUs because the pricing and supply make it worth adapting their workloads, but they can't match the co-design efficiency Google gets by building Gemini and the chip together.
NVIDIA can't go that far because it doesn't own the models. It can't force thousands of customers to redesign their models around its chip, so it has to keep its chips flexible instead of tuning for one model. In fact, a core part of NVIDIA's success historically has been that they built a general-purpose chip that could be used by gamers, crypto miners, and different AI labs.
Jensen used to argue custom chips move too slowly for fast-changing models, and that's clearly not holding up anymore. Google went from 2-3 year TPU cycles, to yearly releases, to shipping 4 different 8th-gen systems just this year (8t, 8i, 8tx, and 8i Max), each tuned for a different job.
Having both the chips and the models is what gives Google a real shot at RSI:
• Before leaving Google, Jeff Dean said AlphaEvolve came up with a circuit going straight into next-gen TPUs, calling it "TPU brains helping design next-generation TPU bodies."
• AlphaChip cuts chip layouts from weeks to hours, and AlphaEvolve sped up part of Gemini's training code by 23% (taking 1% off total training time).
• Amin Vahdat, who runs AI infra, said TPU 8 was co-designed for Gemini, and this week Google showed that building models around TPUs can boost training throughput by up to 57%.
• Sergey said their models are now helping build the next ones, and AI already writes over 75% of DeepMind's production code (with humans still reviewing it).
• Gemini 4 Argon agents also freed up 300+ TB of data center memory on their own and wrote custom TPU kernels running 8x to 15x faster.
• Sundar even told investors on the Q3 call to expect "continued acceleration" over the next 12 months because of RSI.
For most of 2026 Google looked behind and even missed their June Gemini 3.5 Pro target. Then Gemini 4 Argon went live internally, and over the last 7 weeks they've shipped 4 Gemini 4 models (Argon, Carbon, Hydrogen, and Neon).
Neon is only their mid-tier model, yet it beat Opus 5.5 and GPT-6.2 Sol on almost every launch benchmark (only trailing on 3), hit 95% on long context vs 67% and 80%, and cut hallucinations to 14% vs 50%+. On Artificial Analysis it's also about 36% cheaper per task than Opus (Sol is cheaper, though with much higher hallucination rates).
In tech, even a slightly faster rate of improvement compounds into winner-take-most. Right now it looks like Google is improving faster than everyone else.
OpenAI and Anthropic are obviously still serious contenders, especially if eventually scale their own custom chips, though neither has Google's full stack today. NVIDIA was the one company that already had the hardware edge to close this loop from the other side by making huge bets on frontier open-weight models but that is now in question without Jensen there to take those risks. Google here has a real shot at pulling away and in the RSI world if you pull-away, it'll be very difficult for others to catch up as the distance keeps widening.
TPUs are helping Google $GOOG win the AI race by enabling a model-silicon recursive self-improvement (RSI) loop.
Up until recently, RSI was mostly a concept that applied to AI models helping make better AI models. I believe Google is changing that narrative. It's becoming clear that model-silicon co-design is the future of AI, and RSI has a lot more leverage in the model-chip feedback loop.
Because of scaling laws, the AI race has always been more of a compute acquisition race. Model design is important, but secondary to compute. More compute gets you better results, while better models help you extract more intelligence at lower costs. So the race really comes down to first who can accumulate compute at the cheapest price, and then who can use it most efficiently.
TPUs are arguably the biggest leverage Google has in the AI race. Internally, TPUs are 30-40%+ cheaper than comparable $NVIDIA setups, even after paying Broadcom and TSMC, since Google isn't paying NVIDIA's ~75% gross margins. When frontier AI companies are pouring hundreds of billions into compute, that gap means Google gets way more compute per dollar with much better ROIC, and that compounds every cycle.
Selling TPUs to outside buyers like Anthropic $ANT, Meta, and Oracle also lets Google spread its chip R&D across more volume. Because they can afford thinner margins than NVIDIA, they can undercut on price while still keeping the best economics for themselves.
But I'd argue those substantial cost savings aren't even the biggest edge for Google. As I said in my last post, I believe every frontier lab will eventually become a chip company because otherwise you leave too much efficiency on the table.
Because Google knows Gemini's roadmap in advance, they can strip out the extra features a GPU needs and design chips strictly around Gemini's bottlenecks. For example, TPU 8i triples on-chip memory and redesigns chip-to-chip networking specifically for long-context MoE reasoning, giving them up to 80% better performance per dollar than Ironwood.
On the flip side, the Gemini team designs the next model around the TPU's exact layout before the chip is even made. TPUs are much more rigid than GPUs, so outside models designed for NVIDIA can lose half their utilization in key layers on a TPU, while Gemini is tailored to fit it.
Outside buyers like Anthropic and Meta still buy TPUs because the pricing and supply make it worth adapting their workloads, but they can't match the co-design efficiency Google gets by building Gemini and the chip together.
NVIDIA can't go that far because it doesn't own the models. It can't force thousands of customers to redesign their models around its chip, so it has to keep its chips flexible instead of tuning for one model. In fact, a core part of NVIDIA's success historically has been that they built a general-purpose chip that could be used by gamers, crypto miners, and different AI labs.
Jensen used to argue custom chips move too slowly for fast-changing models, and that's clearly not holding up anymore. Google went from 2-3 year TPU cycles, to yearly releases, to shipping 4 different 8th-gen systems just this year (8t, 8i, 8tx, and 8i Max), each tuned for a different job.
Having both the chips and the models is what gives Google a real shot at RSI:
• Before leaving Google, Jeff Dean said AlphaEvolve came up with a circuit going straight into next-gen TPUs, calling it "TPU brains helping design next-generation TPU bodies."
• AlphaChip cuts chip layouts from weeks to hours, and AlphaEvolve sped up part of Gemini's training code by 23% (taking 1% off total training time).
• Amin Vahdat, who runs AI infra, said TPU 8 was co-designed for Gemini, and this week Google showed that building models around TPUs can boost training throughput by up to 57%.
• Sergey said their models are now helping build the next ones, and AI already writes over 75% of DeepMind's production code (with humans still reviewing it).
• Gemini 4 Argon agents also freed up 300+ TB of data center memory on their own and wrote custom TPU kernels running 8x to 15x faster.
• Sundar even told investors on the Q3 call to expect "continued acceleration" over the next 12 months because of RSI.
For most of 2026 Google looked behind and even missed their June Gemini 3.5 Pro target. Then Gemini 4 Argon went live internally, and over the last 7 weeks they've shipped 4 Gemini 4 models (Argon, Carbon, Hydrogen, and Neon).
Neon is only their mid-tier model, yet it beat Opus 5.5 and GPT-6.2 Sol on almost every launch benchmark (only trailing on 3), hit 95% on long context vs 67% and 80%, and cut hallucinations to 14% vs 50%+. On Artificial Analysis it's also about 36% cheaper per task than Opus (Sol is cheaper, though with much higher hallucination rates).
In tech, even a slightly faster rate of improvement compounds into winner-take-most. Right now it looks like Google is improving faster than everyone else.
OpenAI and Anthropic are obviously still serious contenders, especially if eventually scale their own custom chips, though neither has Google's full stack today. NVIDIA was the one company that already had the hardware edge to close this loop from the other side by making huge bets on frontier open-weight models but that is now in question without Jensen there to take those risks. Google here has a real shot at pulling away and in the RSI world if you pull-away, it'll be very difficult for others to catch up as the distance keeps widening.
Up until recently, RSI was mostly a concept that applied to AI models helping make better AI models. I believe Google is changing that narrative. It's becoming clear that model-silicon co-design is the future of AI, and RSI has a lot more leverage in the model-chip feedback loop.
Because of scaling laws, the AI race has always been more of a compute acquisition race. Model design is important, but secondary to compute. More compute gets you better results, while better models help you extract more intelligence at lower costs. So the race really comes down to first who can accumulate compute at the cheapest price, and then who can use it most efficiently.
TPUs are arguably the biggest leverage Google has in the AI race. Internally, TPUs are 30-40%+ cheaper than comparable $NVIDIA setups, even after paying Broadcom and TSMC, since Google isn't paying NVIDIA's ~75% gross margins. When frontier AI companies are pouring hundreds of billions into compute, that gap means Google gets way more compute per dollar with much better ROIC, and that compounds every cycle.
Selling TPUs to outside buyers like Anthropic $ANT, Meta, and Oracle also lets Google spread its chip R&D across more volume. Because they can afford thinner margins than NVIDIA, they can undercut on price while still keeping the best economics for themselves.
But I'd argue those substantial cost savings aren't even the biggest edge for Google. As I said in my last post, I believe every frontier lab will eventually become a chip company because otherwise you leave too much efficiency on the table.
Because Google knows Gemini's roadmap in advance, they can strip out the extra features a GPU needs and design chips strictly around Gemini's bottlenecks. For example, TPU 8i triples on-chip memory and redesigns chip-to-chip networking specifically for long-context MoE reasoning, giving them up to 80% better performance per dollar than Ironwood.
On the flip side, the Gemini team designs the next model around the TPU's exact layout before the chip is even made. TPUs are much more rigid than GPUs, so outside models designed for NVIDIA can lose half their utilization in key layers on a TPU, while Gemini is tailored to fit it.
Outside buyers like Anthropic and Meta still buy TPUs because the pricing and supply make it worth adapting their workloads, but they can't match the co-design efficiency Google gets by building Gemini and the chip together.
NVIDIA can't go that far because it doesn't own the models. It can't force thousands of customers to redesign their models around its chip, so it has to keep its chips flexible instead of tuning for one model. In fact, a core part of NVIDIA's success historically has been that they built a general-purpose chip that could be used by gamers, crypto miners, and different AI labs.
Jensen used to argue custom chips move too slowly for fast-changing models, and that's clearly not holding up anymore. Google went from 2-3 year TPU cycles, to yearly releases, to shipping 4 different 8th-gen systems just this year (8t, 8i, 8tx, and 8i Max), each tuned for a different job.
Having both the chips and the models is what gives Google a real shot at RSI:
• Before leaving Google, Jeff Dean said AlphaEvolve came up with a circuit going straight into next-gen TPUs, calling it "TPU brains helping design next-generation TPU bodies."
• AlphaChip cuts chip layouts from weeks to hours, and AlphaEvolve sped up part of Gemini's training code by 23% (taking 1% off total training time).
• Amin Vahdat, who runs AI infra, said TPU 8 was co-designed for Gemini, and this week Google showed that building models around TPUs can boost training throughput by up to 57%.
• Sergey said their models are now helping build the next ones, and AI already writes over 75% of DeepMind's production code (with humans still reviewing it).
• Gemini 4 Argon agents also freed up 300+ TB of data center memory on their own and wrote custom TPU kernels running 8x to 15x faster.
• Sundar even told investors on the Q3 call to expect "continued acceleration" over the next 12 months because of RSI.
For most of 2026 Google looked behind and even missed their June Gemini 3.5 Pro target. Then Gemini 4 Argon went live internally, and over the last 7 weeks they've shipped 4 Gemini 4 models (Argon, Carbon, Hydrogen, and Neon).
Neon is only their mid-tier model, yet it beat Opus 5.5 and GPT-6.2 Sol on almost every launch benchmark (only trailing on 3), hit 95% on long context vs 67% and 80%, and cut hallucinations to 14% vs 50%+. On Artificial Analysis it's also about 36% cheaper per task than Opus (Sol is cheaper, though with much higher hallucination rates).
In tech, even a slightly faster rate of improvement compounds into winner-take-most. Right now it looks like Google is improving faster than everyone else.
OpenAI and Anthropic are obviously still serious contenders, especially if eventually scale their own custom chips, though neither has Google's full stack today. NVIDIA was the one company that already had the hardware edge to close this loop from the other side by making huge bets on frontier open-weight models but that is now in question without Jensen there to take those risks. Google here has a real shot at pulling away and in the RSI world if you pull-away, it'll be very difficult for others to catch up as the distance keeps widening.
Ok I do have a view on NVIDIA without Jensen.View quoted post on X
I’ll tell you the real reason $NVDA will likely fail without him and why the frontier labs and hyperscalers are the biggest winners. (Elon has been in a really good mood recently. Have you noticed?)
The biggest threat to NVIDIA is its own customers, the frontier labs. NVIDIA has to become a frontier lab faster than frontier labs are becoming chip companies. And with Jensen out, they likely won't maintain the momentum to get there.
With recursive self-improvement (RSI), every frontier lab and hyperscaler must become a chip company or lose the AI race. It's that simple, and they've all laid the groundwork (Google TPUs, OpenAI and Meta building custom ASICs with Broadcom, Anthropic building an in-house silicon team, and SpaceX/xAI/Tesla building Terafab).
That's because software-only RSI has clear physical limits. Software can reduce how much computation and data movement a model needs, but it can't change the chip's physical bandwidth, memory capacity, or interconnect. Custom silicon gives labs another way to improve performance: designing the memory system, compute, and networking around their own workloads.
To get the real compounding effect of RSI, you have to co-design the model and the silicon so it becomes a two-way feedback loop: smarter models design better custom chips, and those chips train and run the next model faster. You can't do that on off-the-shelf NVIDIA GPUs because you get zero proprietary edge buying the same chip as everyone else, and getting your next-gen architecture into NVIDIA's silicon means sharing your roadmap with them years before tape-out (which they then sell to your rivals). Plus, AI now automates writing compilers and porting CUDA kernels to other chips, which kills the CUDA moat.
Jensen argued custom ASICs aren't more cost-efficient than GPUs because models change too fast for fixed hardware, and partners like Broadcom charge fat margins too. But hardware doesn't fall behind model changes when AI co-designs the next model and chip together instead of humans designing them separately (which Google and OpenAI are already doing). And labs aren't building custom silicon just to save on NVIDIA's 75%+ gross margins. Closing that hardware-software loop is the main goal, though the cost savings are still massive since RSI requires 24/7 internal inference where custom ASICs are way cheaper and more power-efficient than general-purpose GPUs.
That's why NVIDIA has been trying so hard to become a frontier open-weight lab, buying Groq (fast inference IP), Poolside + Nemotron (internal coding models), and Hugging Face (distribution). If closed labs win, enterprises just call their APIs running on the labs' own custom silicon and NVIDIA gets cut out. Subsidizing free open-weight models is the only way to make enterprises buy or rent NVIDIA GPUs to run the models themselves.
With Jensen gone, it's unclear if the Hugging Face deal even survives DOJ approval (can a GPU company own the whole stack?), and a non-founder is unlikely to take the massive risks needed to become a frontier lab: putting billions of capex into internal clusters and free open-weight models instead of just selling GPUs, while competing directly with its own biggest customers $GOOG, $META, $SPCX, $ANT, $MSFT
Good thing I listened to Ray Dalio and added uncorrelated assets like gold and bitcoin to my portfolio