
research & investments @polarisfund | exploring web3, ai, energy & frontier tech
Global
Joined September 2015
- Tweets7.4K
- Following6.4K
- Followers3.3K
- Likes22.7K
Imo there will not be such a thing as "enough training data" (atleast for the next 5-7yrs), there will always be need and consumption demand.
Margins will definitely compress i agree, but i also think they won't actually go to zero.
Mainly see an opportunity around them still being relevant for a long period for SLMs/ domain-specific usecases
I feel these data labs will mostly turn into specialised usecases and see exits via acquisitions to one of these frontier AI labs
This quoted post is unavailable.
24
33
1,870
We are also actively exploring n deploying across venture/ liquid/ yield opportunities at @PolarisFund
Pls do feel free to reach out via DMs as well I’d love to just even jam on market research if nothing else :)
I’m also happy to personally give feedback to each n every deck from a researcher pov in the space
I’ll ensure I circle back to everyone n give some concrete feedback to all that do reach out!
Msg me anytime hehe I’m always available
The shift in the crypto fundraising landscape the past 6 months has been insane.
Crypto VCs used to have to constantly be networking/writing/podcasting/going on spaces/promoting your thesis/getting on 10 deal flow calls a week, to get into good deals...now it's literally enough to just have capital to write checks.
Deals are being pushed rather than dug out. Inbound if people know you have money is at an all-time high.
Most firms are either 1) Out of money 2) Moved to Series A and beyond or 3) Fundraising (with no success).
Deals that used to close in 2-3 weeks now close in 2-3 months.
Firms with questionable business models or copy pasta of the latest trend are getting zero primary or follow-on funding (Good news!).
There are now realistically <20 firms writing checks in pre-seed/seed.
VCs basically have the pick of any deal they want, with more time to do DD.
IMHO 25/26 are going to be historic vintages for those who stick around.
I’m also happy to personally give feedback to each n every deck from a researcher pov in the space
I’ll ensure I circle back to everyone n give some concrete feedback to all that do reach out!
Msg me or @PolarisFund - always available hehe
The shift in the crypto fundraising landscape the past 6 months has been insane.
Crypto VCs used to have to constantly be networking/writing/podcasting/going on spaces/promoting your thesis/getting on 10 deal flow calls a week, to get into good deals...now it's literally enough to just have capital to write checks.
Deals are being pushed rather than dug out. Inbound if people know you have money is at an all-time high.
Most firms are either 1) Out of money 2) Moved to Series A and beyond or 3) Fundraising (with no success).
Deals that used to close in 2-3 weeks now close in 2-3 months.
Firms with questionable business models or copy pasta of the latest trend are getting zero primary or follow-on funding (Good news!).
There are now realistically <20 firms writing checks in pre-seed/seed.
VCs basically have the pick of any deal they want, with more time to do DD.
IMHO 25/26 are going to be historic vintages for those who stick around.
8
50
3,226
Was really fun to brainstorm with Jay on this piece, he’s really DIVED into neobanks
Imo its worth a read - I personally feel neobanks in general become one of the core “applications” of crypto which are boring but super necessary in day to day lives of users as we progress ahead.
I spent the last few months writing a 4000 word thesis on "crypto neobanks" - looking at neobank history, market landscape, revenue models, and gaps for builders.
Here's a TLDR:
- The first wave of fintech neobanks (SoFi, Chime, Revolut, Wise) happened in the mobile era. They replaced the frontend of banking from a physical branch to a mobile app. Crypto neobanks are replacing the backend of banking - from traditional rails to permissionless ones.
- Crypto neobanks - like their fintech predecessors - address 4 fundamental interactions with money: Store, Spend, Grow, and Borrow. Incumbents at every layer (Phantom/EtherFi/Binance/Morpho) are racing to consolidate these features into a single "superapp" experience to increase user stickiness.
- The core goal of a crypto neobank is to make money move around faster, using blockchain rails. Some verticals accelerate money more than others, and each has a different monetization model. There's a velocity of money pyramid: Grow > Borrow > Spend > Store.
- Opportunities for builders:
(1) Privacy and compliance parity with existing banks/fintechs
(2) Real world composability (merchant systems, SWIFT, local rails like ACH/Pix)
(3) Leveraging permissionlessness nature of blockchain rails to speed up how money is transferred in the economy, for rich, poor, human, and AIs,
(4) Localization vs. globalization - crypto neobanks can pursue both paths, either with a local-first (Nubank style) model or a global-first model. The former wins through local trust and distribution, the latter through scale and composability.
(5) Undercollateralized lending and consumer credit - this is perhaps the “holy grail” of crypto neobanks, combining compliance/KYC, bridging offchain credit records with onchain ones, and navigating regional differences.
21
56
1,063
🫶
amazing writeup by @DarshanG_ , highly recommend giving him a follow - i've had a ton of fun exchanging notes with him on crypto cards, neobanks, stablecoins and more!
for crypto cards specifically, my current thought is that it really depends on who can drive volumes/user retention to the point where they can capture/dominate the juicy interchange spread - to make this actually a revenue line, rather than free money to rain/reap. perks like better cashback, brand leverage, localization which Darshan writes about are all routes to this goal.
the sad part is that even if actual costs for payments asymptote to 0, the 2.5% interchange may be here to stay, just swallowed up as margins by card issuing protocols. this is actually what happens with walmart/amazon and other retailers that negotiate low rates with card networks and eat that interchange margin.
imo low-fee payments will more likely happen in a novel form factor - QR code/NFC payments rather than through credit cards, which have a legacy rev model.
Have spent a lot of hours looking at 35+ cards and comparing them across distribution, tech, regional n all the other parameters in this long form deep dive, do check it out 🫡
I’d love to discuss in more detail + exchange notes with anyone who would be interested!
It is observed that crypto-cards and neobanks are generally at a great inflection point where now it has become a mainstream category inside crypto where value capture occurs. There are more than 35+ players that we have mapped out, who can be compared with each other across different models and shaped by how they monetise, scale, and acquire/retain users.
Overall there are three value layers:
1. Liquidity owners/partners (defi players) which capture yield + interchange fees
2. Rail owners (infra/B2B) which capture recurring API fees, compliance rent, and treasury/float
3. Distributors/aggregators (CeFi, regional, niche like privacy) which capture interchange/FX but face structural compression.
So, in general, what really makes these card operators defensible or gives them a strong moat? It is majorly around controlling a primitive, which can either be:
(i) regulated access
(ii) programmable liquidity
(iii) deep integrations
In general cashbacks is not anymore a real moat, they are a good to have feature which can help in the initial bootstrapping phase of the product, but doesn’t really become a long-term moat for driving retained users/ loyalty to the platform; but on the other hand providing users rights, access, and reducing switching costs are definitely one of the major drivers.
Let’s understand this well by looking at a few different categories of card operators that have developed in crypto right now:
- custodial exchanges
- web3 wallets
- infra players (more backend)
- regional coverage
- niches like privacy-first
First up, for custodial exchanges or fintech-styled players like Bybit, WireX, and Crypto.com, their major moat as a product lies in the fact that they are able to bring the best of both web2 <> web3 worlds together - they allow you to easily on-off ramp + also buy your fav tokens. So the cards actually become a great way for these players to ‘force lock inside the exchange,’ i.e., make their users sticky by capturing their liquidity and ensuring that they are not moving it elsewhere to other exchanges (more of this).
On the other side, it is also an easy way for users to directly use these cards as a form of onboarding new users to their exchanges (less of this). These players mainly dominate purely on distribution games focusing on licensing, scale, and brand recognition as their major moat, but the economics compress quickly since the interchange and FX fees are capped, and token rewards also fluctuate with market cycles which reduces their profits even if they keep scaling. Overall the success of these exchanges (and eventually their cards) is a function of 1) user churn 2) market liquidity 3) loyalty.
Next there are defi players like Metamask, Etherfi, and Solayer who basically have a dual yield earning structure by internalising both the yield layer (user deposits stETH/eETH/LP positions) and the transaction layer (interchange revenue from spend) which creates two distinct value streams that traditional custodial or exchange-linked cards can’t really capture. This creates a self-reinforcing dual revenue flywheel: more assets → more yield → more rewards → more spending → more interchange → more assets.
It is fair to think that their defensibility can be majorly tied to the idea that users liquidity and usage naturally concentrate where the economics are natively aligned for them. At the same time, there are also a few risks that should be noted about these players since their revenue is highly reliant on 1) defi yields (APY compression, TVL rotation) 2) smart-contract dependencies 3) fiat bridges (operational friction to move assets onchain <> offchain).
The result is that it is a great model with high structural leverage where one can have deeper upside and stronger intrinsic moats, but volatility does still exist and is heavily influenced by market conditions and protocol health.
Infra providers like OrbitX and Kolo.in actually operate at a completely different layer of the stack. They are basically the ‘backend infra’ rail for many of these other companies who actually build frontend experiences on top of them and compete for acquiring new users, card designs, or cashbacks. What is handled by these infra players are aspects like issuing cards, moving money, converting between fiat and crypto, and staying compliant with regulators. Their business model is built around integrations with partner projects via APIs, SaaS fees, FX/settlement rails, and compliance tooling. The best way to think about them is that whenever a partner project performs an action for a user like issuing a card, topping up a balance, performing KYC, or running an FX conversion, it is actually handled by these players and they take a small cut.
Overall, it is fair to think that once many integrations are through, the revenue streams are somewhat recurring and predictable + tied to a focus on ecosystem growth overall rather than singular user metrics like AVPU. It is actually a pain for a company to switch since they would have to rebuild their card logic, re-establish compliance channels, redo integrations with banks/networks, and migrate users, which who would want to do, right?.
This is what works well for them (ofc if done well), creating deep, compounding stickiness over time. They are very comparable to web2 payment processors or custody infra tools which are barely front-user facing but actually do all the heavy lifting on the backend. It is more of a tech moat (ensuring the rails work + easy to integrate) + sales (landing good B2B clients which ideally would have good distribution and ensuring they don’t churn).
Regional players like Lemon, Kasi Money, and Sora have actually built their entire value around operating deeply inside a specific (localised) market. Their strength comes from being able to have local payment rails, fiat–crypto conversion paths, domestic banking relationships, and regulatory clarity.
In this case the moat is majorly a combination of understanding local user behaviour + ensuring the infra is highly fluid in the native localised setup with different partners. But ofc, localisation has its own upside limitations because it gets tough to start easily scaling this model in a new country, as one would need to again figure out new licences, forming new banking relationships, adjusting to different consumer expectations, and often redesigning the onboarding completely. So overall these players are regional masters, but have a hard time globally, making expansion slow, expensive, and operationally heavy.
Lastly, there are cards focused on a niche like privacy-focused cards such as Payy whose entire value prop is built around anonymity, censorship-resistance, and minimal disclosure, giving them a clear narrative and a loyal niche audience. Users choose them because they are more ‘ideologically aligned than purely the technological infra’. Generally, creating a brand and providing a more non-invasive experience are their major focuses.
One thing these players majorly face issues with is that in regions requiring strict KYC, AML, and monitoring, compliance becomes a huge hurdle for these players to penetrate, especially when looking at more offchain transactions. Compared to all the other categories, privacy for eg: makes a very small set of user cohorts who are highly ‘privacy-first’ aligned (more ethos/ principles based than purely product experience).
Just to summarise the learnings a few objective takeaways are:
- value is shifting down the stack where real moats now come from infrastructure, compliance, and liquidity control rather than rewards or branding
- regulation is becoming the main determinant of scale, with licensing and bank access deciding who can expand into new markets/geographies
- owning the liquidity layer helps to drive revenues, as one can control yields and balances rather than competing over thin margins
- integration depth compounds defensibility
- local advantage still matters, where corridor-specific rails, FX pathways, and domestic partnerships can outperform broader players
- cefi cards deliver high volume but low leverage, remaining dependent on subsidies and thin FX/interchange economics
- defi cards offer stronger but more volatile economics
- infra rail providers once integrated are super sticky
- niches like privacy-led are more focused on narratives/branding than purely on product moats as compared to other categories, which leads to more friction in adoption + scaling
So, what would really make sense doing here?
- owning a primitive like licensing coverage, integration rails, or liquidity pathways
- creating stronger models to monetise flows around inevitabilities (volume, balances, compliance events) rather than depending on yield that comes from token inflation
- designing for a way where an operator is irreplaceable in the flow via multi-party integrations, regulator attestations, custom risk engines, and issuer–processor dependencies that can create natural switching barriers
- optimizing the flow from on-chain → off-chain by offering instant liquidity backed by crypto collateral, reducing user friction and strengthening issuer risk controls
Lastly, a few whitespaces or opportunities we identify are around:
- compliance orchestration: unifying KYC/KYB, travel-rule, sanctions, and audits into one API; the moat here comes through regulator relationships, approval data, and workflow entrenchment
- smarter FX & routing: basically a routing aggregator across CEX/DEX/RFQ with guarantees and treasury backstops; the idea is to monetise micro-fees per routed transaction + spread share
- credit on onchain collateral: real revolving credit against stables/LSTs with LTV, liquidation buffers, and auto-repay; the moat is in having highest ARPU and having a setup of a good oracle network, liquidation logic, and risk engine
- regional issuer kits: prepackaged UPI <> SEPA / PIX <>USD stacks (BIN, KYB, FX, settlement, tax) acting as ‘regional issuer-in-a-box’; the moat comes from sponsor-bank networks, local licences, BIN scarcity, and regulatory support (which are all super tough to acquire)
- legal privacy: selective disclosure (ZKP/ TEE) that preserves user privacy while remaining regulator-grade; the moat is built through policy-backed attestations, trust marketplaces, and regulator comfort
- yield-aware settlement & float management: a ledger that routes idle balances into approved yield, unwinds instantly for spend, and stays auditor-friendly; the moat is enabling regulated, risk-controlled yield across CeFi and DeFi
Looking at it generally, one of the bigger opportunities right now could be something around building a compliance + FX + issuance middleware layer with optional yield and credit modules. Whoever becomes the Stripe/Checkout.com for crypto cards along with real regulatory cover will ideally end up owning the entire category (full-stack approach).
Closing in, the real value is no longer a unidirectional mechanism (who issues the card/ who offer the biggest cashback), but essentially it is a combination of a few aspects like controlling the underlying primitives around licensing, liquidity routing, compliance logic, and the settlement rails.
Overall it feels like it is about being able to identify where the real leverage and margin exist and ensuring these players are tapping into the opportunity.
there have been so many cards that have been popping off lately, have been diving deeper into this topic and feel it's a great unlock which can actually help add net new users/ use-cases to crypto
do check out this deep dive focusing on the diff types of cards flows that we believe exist currently 👇
Crypto cards are at a pretty interesting inflection point in my opinion.
They’ve quietly bridged the old and new payment worlds, letting you pay from a crypto wallet while the merchant still receives fiat through the same Visa or Mastercard rails they already use. To the shopper and the store, it feels like a regular card payment; the crypto part just happens behind the scenes.
What’s really changing is how the payment is settled on the backend - when and how value moves between crypto and fiat, rather than the front-end experience.
To understand what’s actually changing in the background, it helps to look at how these cards are structured today.
There are two broad setups that define how crypto cards work behind the scenes.
Prepaid cards convert your crypto or stablecoin into fiat before spending. You top-up a fiat balance first, and every transaction simply draws from that prefunded account, it's almost always debit.
On the other hand, real-time cards keep the funds in your crypto wallet until the moment of purchase. When you tap your card, the system checks your wallet, converts just enough into fiat, and approves the transaction instantly. This model can work for both debit and credit programs.
Every card payment still follows two clocks:
- a fast “approval” so you can walk away
- a slower “settlement” when actual money moves between the merchant’s provider, the network, and the issuer (usually the next business day or two).
Over time, we believe that some of these flows may eventually settle directly onchain in seconds, with an optional off-ramp to bank money.
In today's setup, the fee stack looks like the following:
→ merchant’s provider takes a service charge
→ issuers earn interchange
→ networks charge scheme fees
→ wallets or APIs earn a conversion spread when crypto becomes fiat.
The biggest near-term change is in business and cross-border payments, where stablecoins could potentially cut down the steps in the flow and help shorten settlement windows.
Each of these setups plays out differently when a payment actually happens: the difference lies in what gets converted, and when.
We’ll dive further into the crypto cards ecosystem in future blogs, but for now let’s look at each of the flows in this high-level chart.
→ Prepaid flow (prefunded fiat; debit only)
In this setup, the user converts crypto or stablecoins into fiat before spending. The issuer approves payments against that prefunded fiat balance, and the merchant later receives fiat through normal card settlement.
a) before tap: convert crypto/stablecoin to fiat and load the card account
b) path: POS terminal → acquirer/PSP → Visa/Mastercard → issuer
c) issuer action: check prepaid fiat, place hold, authorize
d) settlement: via card rails (T+1–T+3)
→ Real-time debit flow (linked wallet/API)
Unlike the prepaid model, here the user’s funds stay in their crypto wallet until the moment of purchase. At authorization, the program checks the wallet, locks the amount, converts it to fiat in real time, and approves the transaction. The merchant still receives fiat later through standard card settlement even though both sides can see the approval instantly.
a) path: POS terminal → acquirer/PSP → Visa/Mastercard → issuer/program
b) program API: check wallet → lock → convert to fiat → authorize
c) issuer approves based on that conversion and lock
d) visibility: instant (webhooks); settlement remains T+1–T+3 on card rails
→ Real-time credit flow (credit line; wallet-referenced optional)
This is the credit variant of the real-time model. Here, the issuer approves the transaction against a credit line sometimes also referencing the user’s wallet balance as a risk signal or even collateral. The merchant still receives fiat later through standard card-network settlement, with all the usual credit-card economics at play.
a) path: POS terminal → acquirer/PSP → Visa/Mastercard → issuer
b) issuer action: check credit line (wallet may inform risk) → authorize
c) settlement: standard card clearing (T+1–T+3); credit economics apply
→ Onchain settlement flow (future / limited pilots)
The next evolution removes card-network clearing altogether. Here, the payment is sent directly to a blockchain by the POS or payment gateway, and the merchant receives funds within seconds in their onchain wallet. There’s no Visa or Mastercard step for that transaction, it settles natively onchain. The merchant can off-ramp to bank money anytime if they prefer fiat.
a) path: POS/gateway → L1/L2 → merchant wallet (no card-network clearing for that payment)
b) outcome: seconds to finality, optional off-ramp, programmable rewards/tax/splits in-transaction
Now that we’ve mapped how each model works, the next question is which parts of this system will change first and where the biggest breakdowns might occur.
Where will the disruption really happen in our opinion? Key takeaways.
→ B2B(bank to bank) get disrupted first
While consumer card programs are the visible layer, the earliest disruption likely comes in b2b and cross-border flows. These are still handled through backend bank connections and SWIFT messaging which are slow, opaque, and reconciliation-heavy. Moving them onchain would enable faster finality, fewer breaks, and cleaner treasury positions.
→ Elimination of interbank card networks (visa/mastercard)
In the current setup, Visa and Mastercard behave like the interbank networks that almost everyone uses every day, and most crypto card payments still rely on them for clearing and settlement. If a payment is handled fully onchain end-to-end, that specific payment does not use card-network clearing and can feel like a direct bank-to-bank experience for both the user and the merchant with potentially eliminating the need for interbank operators like Visa or Mastercard for that flow.
→ Settlement finance (the T+1/T+2 gap get's bridged)
Traditional settlements take one to two days (T+1/T+2), so banks offer short-term financing to bridge that gap. If the payment leg settles onchain with finality in seconds, that gap disappears and the need for settlement-gap financing goes away.
→ Merchant funding time reduces
For merchants today, even when a crypto card approves instantly, the actual funds still arrive through card settlement a day or two later. If that same merchant (or its gateway) accepts an onchain payment for that transaction, the funds can arrive in seconds and card-network clearing don't be a part of that transaction.
→ Product axis changes (prepaid vs real-time)
For card programs, the real axis is when crypto turns into fiat: prepaid converts before spending, while real-time converts at the moment of authorization. Custody mainly changes how the lock and conversion are done (an exchange hold versus a smart-contract allowance) rather than the path through the merchant, acquirer, network, and issuer.
→ Two clocks in the same payment cycle
Every card payment follows two clocks: the “yes/no now” decision happens in seconds so the shopper can leave, and the “money later” movement happens after clearing so the merchant gets funded. A fully onchain payment can compress both steps into seconds for that specific transaction.
→ Value accrual and economics capture shifts
Economics follow control of conversion and settlement. Today issuers earn interchange, networks earn scheme fees, acquirers earn the merchant fee, and wallets or APIs earn a conversion spread when they convert crypto to fiat. As more flows run onchain, a larger share of value and data shifts toward whoever operates the onchain rail for that corridor, while a smaller share remains in card-specific fees.
All of these shifts point toward one simple reality that the experience at the front may stay familiar, but the rails beneath are the ones that'll change.
To summarise, in the near term, B2B and cross-border payments could see the biggest wins or disruptions: faster funding, fewer reconciliation breaks, and cleaner treasury positions.
Overall, there’s a really strong market potential, and we believe that crypto card and neobank models can, in general, deliver a better user experience than the existing web2 solutions.
Excited to dive deeper and compare the various cards and their offerings over the next few reports.
Spent a lot of time diving into the economics behind the CEX/ DEXs token buybacks/ burns and especially how they are a big contributing factor to their platform economics
It's interesting to see almost all the exchanges have been doing it consistently yoy over the last few years which is via:
- % of their profits/ revenues (only buyback or buyback + burn)
- directly from their treasuries
Do check out the detailed research, keen to speak more with anyone spending more time across this verticals and exchange notes 🫡
The Economics of CEX Token Burns 🔥
Buybacks and burns by exchanges are not new
They have been running quietly for years, shaping supply and demand long before they drew mainstream attention
Almost every large CEX including Binance ($BNB), OKX ($OKB), Gate ($GT), KuCoin ($KCS), and MEXC ($MX) has operated some form of burn program for more than half a decade
What has changed is how they are framed and marketed.
$HYPE (Hyperliquid) put buybacks at the center of its token strategy, not buried in the fine print. It turned burns into a defining feature rather than a background mechanic
More importantly, Hyperliquid runs them consistently and openly as part of the treasury process, setting a new benchmark for clarity
That positioning made burns look novel, even though incumbents like Binance, OKX, Gate, KuCoin, and MEXC had been doing the same for years
The difference is they never marketed them as aggressively or integrated them as tightly into treasury operations (my thoughts on why did they so below)
Burns function as mechanisms of value transfer. They show:
- how an exchange connects token supply to its business model
- which levers drive scarcity (profits, formulas, or governance)
- how credibility is built or lost over time
Burns also act as inflation controls, offsetting token unlocks or emissions and stabilising supply over time
The question now is no longer if burns will happen, but how consistently they are executed, and whether the model provides clarity to tokenholders
To see the shift clearly, here’s a breakdown of how supply dynamics play out across exchange tokens
🧠Takeaway: When analyzing an exchange token, the burn model matters. Profit-based, formula-driven, and governance-controlled designs create very different outcomes for scarcity, predictability, and trust
⚙️ Burn models as operating systems
Exchange burn programs fall into three categories:
→ profit or revenue linked (Gate, KuCoin, MEXC): a fixed % of earnings is used to buy and burn tokens. Cadence is predictable and auditable
→ formula or treasury driven (Binance, OKX, Bitget): supply cuts are set by formulas or treasury allocations. Scale is larger, but the tie to business health is weaker
→ governance driven (Bybit, HTX): cadence is decided through tokenholder votes. This decentralises control but introduces political and execution risk
Programs do evolve. For context: Binance it moved from profit-linked burns to a formula based on price and block counts, and later added BEP-95 gas fee burns. At points it changed the burn rate suddenly, presenting it as proof that BNB was not a security. From a regulatory view this made sense, since decoupling burns from profits lowers classification risk, but the shifting mechanism left markets uncertain
Additionally, other CEXs also updated their dynamics, like:
- KuCoin tightened cadence to monthly for clarity
- Gate has kept its 20% profit-based allocation steady since 2019
With fixed max supply tokens, burns are less common. When they do occur, they’re valuable because they reduce circulating supply and accelerate the path to full supply compression
🧠Takeaway: The burn model decides durability
→ Profit-linked = steady and auditable
→ Formula = scalable but opaque
→ Governance = decentralised but harder to rely on
→ Sudden model shifts (Binance) create structural risk, while transparency moves (KuCoin) build trust
⚖️ Regulatory Angle
Burn models aren’t just economics. They’re also regulatory positioning
In traditional equity markets, corporate buybacks have always been controversial. The SEC questions them for:
- market manipulation
- insider benefit
- weak disclosure standards
Token burns are crypto’s version of buybacks, but they run without those legal guardrails. That gap changes how models are designed.
→ Profit-linked burns look most like buybacks. They invite closer regulatory attention, since they tie token value directly to profits.
→ Formula-driven burns (Binance’s auto-burn, OKX’s supply cap) are easier to defend. They can be presented as mechanical, detached from revenue, and less likely to trigger securities classification.
→ Governance-driven burns add politics. Regulators may see community votes as insufficient safeguards against manipulation
🧠Takeaway: Burn design is part tokenomics, part legal defence. Decoupling burns from profits reduces regulatory risk, but it also reduces clarity for tokenholders
📊Patterns across exchanges
Three clear patterns stand out:
→ scale with opacity
- Binance burns the most (~$1B/quarter) but keeps shifting rules
- OKX eventually capped supply at 21M after years of cadence
Takeaway: scale commands attention, but rule changes and delayed caps dilute clarity
→ steady profit-linked cadence
- Gate: fixed 20% of profits since 2019, ~60% supply removed
- MEXC: 40% of profits, ~57% cut so far
- KuCoin: moved to monthly cadence in 2022, but burn size shrinks with profitability (10% of profits)
Takeaway: profit-linked models are most forecastable. Smaller burns = weaker business health
→ new entrants and governance risks
- Bitget: $5B burn in Dec 2024, now ~30M tokens/quarter, aiming for 95% cut
- Mantle: burned 98.6% of BIT in migration; now DAO-dependent
Takeaway: headlines do help, but only proven cadence builds durability
⏱️ Cadence, size, recurring burn quality
Supply cuts vary from ~30% at Binance to ~93% at OKX. But markets don’t price percentages alone. They price recurring burn size and the predictability of cadence.
a) Gate, KuCoin, and MEXC: steady profit-linked burns → trust
b) Binance: biggest numbers → blurred by constant formula tweaks
c) OKX: reinforced confidence with $7.6B mega burn → backed by years of consistency
d) Bitget: $5B burn → unproven first event, need to see how it evolves
e) Cryptocom: reduced trust by reversing its 2021 burn in 2025
Note: The August 2025 mega burn coincided with a sharp rally in OKB, showing how one-off supply events can drive near-term price action at times
🧠Takeaway: I think one should don’t just track tokens burned. Rather we should ask: is scale repeatable? is cadence tied to profits? is governance stable? Focus should be on recurring USD burn per quarter/year. Mega burns = credibility markers, not guarantees
👀 Our overall view is that
→ Consistency matters more than size. Markets reward repeatability for sure over making headlines
→ Profit-linked models are best-in-class. They tie token value to exchange health and are easy to evaluate (transparent)
→ Mega burns are only markers. Without follow-through, they fade into optics
Overall it seems that still buybacks are a major marketing expense for now
But what's interesting is that implicitly, CEXs are choosing to recycle profits into their own tokens rather than hold retained earnings in cash or USDC
This concentrates treasury value in the token itself, magnifying both upside and risk at the same time
Finally, as we tried to cover in the overall report as well, the real innovation is not in burns themselves, but in how consistently and transparently they are run.
Hyperliquid definitely has reframed burns as a visible, recurring treasury function
That has actually reset expectations across the sector: scarcity alone is not enough now.
Now, scarcity that is regular, clear, and aligned with exchange economics is what is super valuable now for all the exchanges, which changes the dynamics for many of these CEXs who very slow to move
It'll be interesting to track how this evolves over a period of time.
Put down a detailed comparative analysis between different robotic foundational models.
Check it out 💪🏼
Large language models improve in a predictable way. More text, more parameters, and more compute mostly lead to steady improvements in performance.
But RFMs don’t really follow the same pattern. As we discussed in the last report as well, their progress + progression is highly uneven and shaped not just by "data" but also by many external factors such as body of the robot, the sensors it carries, and the type of data used in training.
Overall, the RFMs can be segregated into different groups basis really the type of task, movement and navigation that’s required for the specific use-case. On a high level (ofc this is a non-exhaustive list), but here are a few broad categories I’ve seen:
→ Generalists: RT-1, RT-2, RT-X, PaLM-E, Gato (these are built to cover a wide range of tasks and robot types with images and language)
→ Manipulation: RFM-1, π0.5, ManipLLM, ManiFoundation, DYNA-1 (they are focused on handling objects such as picking, placing, gripping, and fine control)
→ Navigation: ViNT, NoMaD, PACT, CroCo v2 (they are built for robots that need to move through new spaces, reach targets, or explore unknown terrains)
→ Humanoid/coordination: GR00T N1.5, Helix, Skild Brain, Humanoid World Models (these are aimed at whole-body motion with many diff joints working together)
→ Hybrids: Gemini Robotics (these combine planning with direct action,using simple coded steps to guide tasks)
In the table below, I’ve just really tried to map/ compare these models side-by-side. Do take a look (full post on website as well)👇🏻
Some high level observations from the table are:
- model size ranges from very small to extremely large, but bigger models do not necessarily always deliver better results
- inputs may vary across models, with mostly multi-modal inputs like many using only images and language/ text while others also include signals such as action history, robot state, touch, or LiDAR
- outputs for each of the cases are handled differently, as some models generate step-by-step commands while others produce continuous motions/ full action sequences for navigation
- the design of each model reflects the type of robot it is built for generalists are trained across many robots
manipulation models focus on object handling
navigation models specialise in movement
humanoid models manage whole-body control
hybrids mix these approaches
- each model is defined by a key feature, such as cross-robot training, warehouse data, reasoning traces, contact synthesis, or code-based control, and these kind of shape their strengths and limitations
- data quality often proves more important than dataset size, since smaller datasets with detailed contact and failure cases are always far more valuable than generalised large datasets
- evaluation remains a challenge because there is no single benchmark, meaning that most models still depend on real-world testing to show reliability and versatility
Overall, I feel robotics does not really yet have that “universal foundation model” which works well for all use-cases as each task demands a different data point.
I think this will be the final post in the robotics series for now. Over the last few months, we’ve tried to dive deep into different aspects of robotics from the stack, to the bill of materials, to the flywheel and how they really together.
Thanks for following along, and I’m excited for what comes next!
Tried to put down a quick comparative analysis between the issues that RFMs face over traditional LLMs 👇
With language models, scale is simple: feed them more text, bigger networks, and more compute, and they usually get better.
Robotic foundation models (RFMs) are built a bit different. They live and operate in the physical world with hardware, sensors, moving parts, and real human contact.
The result is a red curve that jumps and dips instead of climbing smoothly (in case of LLMs) as you can see above.
A big reason is that every dataset is tied to a particular robot or more precisely, to specific use cases (as we explored in the last report), such as a warehouse arm, a quadruped, a drone, or a humanoid. They all move, sense, and act in their own ways.
Without a shared “language” for actions, data from one robot often can’t be used by another. For example, a grip from a parallel gripper doesn’t translate well to a five-fingered hand.
Robots also take in many kinds of inputs: cameras, depth sensors, touch pads, joint encoders, microphones, and language instructions.
Each type of sensor works on its own clock like security cameras and microphones which record at different frame rates. If they aren’t in sync, the robot ends up with an incomplete picture, sometimes worse than not having that input at all.
Data diversity > quantity. For eg. in a search engine, crawling a billion copies of the same webpage doesn’t improve results. What matters is covering as many unique sites as possible.
Similarly in robotics: repetitive free-space motion logs won’t help as much as varied actions that include contact (where the robot is actually touching and interacting with an object), edge cases, and different sensor inputs.
There are major three tiers of the quality of data. They are:
→ Top tier data can be sourced from channels like:
- directly collected robot sensor data eg. a robot arm logging its exact grip force while packing boxes
- expert demos with tactile/force data
- realistic verified simulations
→ Mid tier data comes from:
- human video (in first person) adapted to robot actions eg. GoPro footage of a chef cooking, converted into steps a kitchen robot can follow
- play data re-labeled for object affordances and skills
→ Low tier data majorly includes:
- long motion logs with no goals/contact eg. hours of a drone flying around without touching or interacting with anything
- old 2010s datasets which add very little to modern generalised models
RFMs also have to deal with something language models don’t: turning a simple plan into smooth, real-world movement. For example, saying “pick up the cup” is easy, but actually doing it means making dozens of tiny adjustments in grip and position. Even being off by a few millimetres can make the cup slip and that can be costly.
Testing is hard as well. Language models can be easily be benchmarked in minutes.
RFMs actually need thousands of physical trials, which are slow, messy, and expensive. Simulations can definitely help narrow down the choices, but they can't fully capture things like how slippery a surface is or account for those rare failures that might show up once in a hundred runs.
Grouping tests by challenge, like new objects, strange lighting, different tools, or unfamiliar robots can help spot problems earlier.
I think it's also worth understanding when does scaling helps vs hurts👇
Helps when:
- robots move the same way (eg: similar warehouse arms)
- data has real contact + mistakes (eg: drops, slips)
- sensors line up in time (eg: video + touch match)
- training mixes methods + clear steps (eg: find → reach → grip)
- prompts use pics, arrows, or skill names
Hurts when:
- robots are too different (eg: legs vs wheels)
- data in "free-space motion" (eg: drone in empty field)
- touch data missing or out of sync
- no step breakdown (eg: “set table” as one task)
- data skewed (eg: too many apple picks)
- only text prompts for hand work
- no feedback to fix repeat mistakes
So, what can really help move the red curve upwards? Making it more smooth.
- more context bandwidth: motion sketches, hints on object use, skills you can steer
- pre-train → post-train: start broad, then fine-tune for specific jobs
- cheaper & standard hardware: developments from players like Unitree, Figure, Optimus
- better evaluation: test on real gaps, simpler rankings, live metrics
- fleet ops + teleop: keep the data loop running with human assist when needed
Let's probably understand a few of them more deeply👇
→ Context bandwidth
- motion prompts like drawing a line on a map for a delivery driver; the robot sees the path it should follow and can try again without “relearning” the whole route.
- affordance prompts like showing someone where to hold a shopping bag so it doesn’t tear; the robot gets hints on where or how to grip, making it steadier in new situations.
- skill prompts like giving a chef a list of recipes they already know; the robot reuses skills such as “pick up cup” or “turn knob” and combines them for new tasks.
→ Training phases
Think of RFM training as recurring loop across these three major stages:-
- pre-training broadly on diverse sources such as web images and text, actions from multiple robots, human first-person video, and realistic simulation (e.g: shapes from online pics, moves from warehouse arms)
- post-training narrowly with short demos on the target robot, remote corrections, and fine-tuning from human feedback (e.g. shelf picking, steering grip, rating gentle handling)
- after deployment working on logging failures, recreating tricky scenarios in simulation, and learning from moments when a human takes over (e.g. cup drops, dark rooms, human saves)
The takeaway
In robotics, RFMs deal with problems that LLMs don’t. Making models bigger or adding more data isn’t enough. Differences in hardware, gaps in supply, and unpredictable real-world conditions can still cause drops in performance.
For steady progress, focus on hardware, operations, and data feedback as much as on the model itself. Use standard parts, have backup suppliers, and feed real-world data back into training for eg: like a F1 team that adjusts the car, replaces parts, and reviews race data after every run.
Over time, this can make performance more stable and reliable.
Tried to zoom out and answer a deceptively simple question:
How do different types of robots really compare across tasks, environments, and capabilities?
Not just “what can it do?” but “what tradeoffs are each of them optimized for?”
This quick visual breakdown looks at:
- how well robots perform across mobility, sensing, autonomy, etc
- where they fit best across use cases like warehousing, inspection, hospitality, etc
- why form alone is no longer the defining factor - integration is
tldr: The shape matters less than the system-level fit
do check out the heat-maps & entire breakdown below across diff robot types 👇
Robots today come in all forms - walking, rolling, flying, swimming, even squishing through tight spaces.
Their diversity isn’t just about how they look: it’s about what they’re built to do. From humanoids that mirror human motion to drones that chart the skies, each type reflects tradeoffs in environment, mobility, and use case.
Some are generalists like humanoids are built to take on a wide range of physical tasks across industries.
Others are specialists: robotic arms deliver precision but stay fixed; drones cover vast areas but can’t carry much; swarm bots are weak alone but powerful in numbers. Quadrupeds bring stability across uneven terrain. Wheeled AMRs thrive in fast, structured spaces. Underwater robots explore where humans can’t. Soft robots handle what rigid ones can’t like fruit, tissue, fabric.
Zooming out, the landscape spans (non-exhaustive list):
• humanoids (eg: Tesla Optimus, Figure 01)
• quadrupeds (eg: Unitree Go series)
• wheeled robots & AMRs (eg: Kiva, Locus, OTTO)
• robotic arms (eg: Fanuc, UR5, Da Vinci)
• service robots (eg: Pepper, Temi, PuduBot)
• drones (eg: DJI, Zipline, Skydio)
• underwater bots (eg: BlueROV2, Saab Seaeye Falcon)
• soft robots (eg: Soft Robotics, MIT research bots)
• swarm systems (eg: Kilobots)
service robots ofc could ideally have overlaps with other categories
Each form balances autonomy, mobility, and interaction differently; shaped as much by the physical world as the digital logic inside.
What defines a robot’s role isn’t just its shape, but also
• it’s its sensor stack
• its interaction model
• the kind of work it’s built to do
Bots made for labor tend to be stronger and more rugged. Bots made for service are lighter, faster, more aware. In domains like healthcare, hospitality, or warehousing, the real constraint isn’t ability, but it’s actually context: how well the robot fits alongside humans, not just how well it performs.
What’s emerging is a toolbox of robot types: each tuned for different terrain, task, or interaction style.
Understanding these tradeoffs say like movement vs. manipulation, precision vs. adaptability, solo power vs. coordinated scale is key to deploying the right form for the right job. A drone and a humanoid might both “move,” but the cost, complexity, and upside of that movement are entirely different.
What matters isn’t the shape, it’s the system-level fit: how a robot senses, decides, and interacts in its operating environment.
As lines blur between types, form alone says less and less. The edge now lies in integration: pairing the right mobility, sensing, compute, and autonomy stack to solve for a specific constraint or opportunity.
The real question imo one should be asking isn’t “what kind of robot is this?” It’s “what tradeoffs is it optimized for?”
6
18
808
Put down a short report on robotics flywheel and how each of the components link with each other 👇
Robotics has historically focused on building better hardware and tightly engineered control systems.
But the real long-term advantage is shifting toward something far less visible 'proprietary motion data' - which is generated by fleets of operating robots.
Motion logs capture joint angles, torque values, sensor inputs, and success/failure signals which are the raw data that powers learning and model refinement over time. It’s this kind of data that enables foundation models to improve and generalize over time.
In today's world, a good model on its own isn’t enough anymore. For context, even models like GR00T, Helix and other upcoming VLA models will likely become accessible or replicable overtime.
So, what sets systems apart is the feedback loop grounded in real-world experience, a loop that forms a flywheel:
→ deploy robots (even if they’re not perfect)
→ collect motion logs over millions of hours
→ train and fine-tune models on that data
→ improve performance in the field
→ drive more usage and deployments
→ repeat
The more robots you deploy, the more edge cases you encounter. And with every bit of motion history you capture, your system gets smarter. That loop compounds and the moat deepens with each cycle.
In my opinion, real defensibility will now come from:
→ telemetry across diverse environments
→ hundreds of millions of task attempts: picks, failures, recoveries
→ logs labeled with failure types, outcomes, and context
→ model updates tied to measurable performance gains
→ higher success rates on edge cases: low light, occlusion, unstable terrain
Simulation tools like Isaac Sim and Mujoco are useful for early-stage learning. But they can’t necessarily replicate the messiness of the what a real physical world provides: slippery floors, unpredictable humans, crushed packaging, broken lighting and much more.
Real-world deployment creates edge-case data you can’t fake.
This is where full-stack players have a real edge. When you control the hardware, the software, and the data, you're not just learning faster, but you're also building tighter systems. And those systems evolve faster with every deployment.
Thus players like Tesla, NVIDIA, and Figure are not just building robots they are also controlling:
→ the physical systems
→ the motion logs
→ the foundation model
→ the orchestration layer that adapts on the fly
→ the deployment surface: where learning actually happens
→ the feedback loop that ties it all together
Motion logs are the new model weights.
The next wave of robotics leaders will be those who capture most diverse, high-volume, high-resolution telemetry, and build systems that continuously learn from it.
11
18
711
Bill of materials is one of the hidden strategy components to study about a robot.
I've dived deep into it - helping break down why it not only matters, but also helps you map out the stack + figure out always where's the next opportunity going to come through
Hope you'll find it cool 👀
Humanoid-robot economics look complex. Until you zoom into the Bill of Materials (BOM).
BOM isn’t just a cost sheet. It’s a "strategy map".
It reveals where cost hides leverage, where complexity hides moat, and where hardware and software meet or clash.
- when you know how to read it, the BOM tells you:
- which subsystems drive cost vs. differentiation
- where integration complexity creates or erodes margin
- what IP is worth owning (mechanical, model, or data)
- how geography impacts iteration speed
- why some companies scale like SaaS and others get stuck in hardware
Let’s walk through the robot stack with the BOM lens on🔍
⚙️Actuation drives cost and moat
At the base of every robot is actuation. That’s where nearly half the BOM cost sits and where most of the defensibility lives today.
Motors, gearboxes, torque systems:
- have complex supply chains
- require high levels of integration
- are extremely competitive
But they’re also strategic strongholds.
Vendors like LeaderDrive, Harmonic Drive, and Beite used to dominate here with hard-to-copy IP built on:
- tight tolerances
- decades of iteration
- end-to-end vertical integration
If you control actuator design, you control a robot’s core physics.
These systems though are now getting replaced AI centric ones, which favors on dynamic behavior and cost, being less IP centric and hedge more on strong manufacturing and good overall product design.
🧠 Takeaway: high cost + high IP = hardware moat
💻 Software eats margin
But moving up the stack, the leverage flips.
Compute, sensing, and model layers might each take up ~10% of the BOM, but they disproportionately define:
- adaptability
- autonomy
- cost compression over time
Chips are cheap. Torque isn’t. But paired with the right stack, compute compounds fast.
That’s why companies like NVIDIA don’t stop at silicon. They own:
- Isaac → simulation tooling
- GR00T → foundational control models
- Fleet learning loops → systems that get smarter with usage
This is the GR00T playbook: Own the model, the data, and the learning loop and suddenly, your 10% of the BOM controls 80% of the product’s differentiation.
This is where margin scales like code:
- fleet-level learning loops
- API-style unit economics
- lower per-unit cost over time
usage → data → model → autonomy → usage → margin
🧠 Takeaway: low cost + model/data IP = software moat
🧩 Integration creates margin or kills it
Hardware margins are brutal - unless you own precision complexity.
The question is: are you assembling parts? Or engineering behavior?
A company might carry high COGS if it:
- makes its own actuators
- tunes torque curves for specific behaviors
- designs electromechanical systems end-to-end
But that’s not just about defensibility, it’s a risk-for-margin tradeoff.
Owning this layer means taking on supply chain complexity, firmware tuning, and precision control. But if done right, it can knock 25–50% off costs which is a serious edge for hardware-heavy players.
Case study: Unitree’s $16K humanoid
Unitree didn’t win only on AI breakthroughs. They won by engineering the BOM:
- sourcing actuators locally (cheaper motors) + in house manufactured (no vendor margins)
- skipping multi-plane LiDAR (cutting sensor cost)
- avoiding radical architecture; just tight control over the stack
- co-locating supply chain; 90% of vendors within hours of Hangzhou
Here's a high level breakdown of the BOM for Unitree's Quadruped:
Western companies often can’t match this - not because of tech gaps, but because they lack supply chain density.
That slows iteration and bloats the BOM.
🧠 Takeaway: low BOM + tight integration = speed + margin
💸 Cost ≠ Commodity
A low BOM doesn’t guarantee margin.
A high BOM doesn’t mean weak strategy.
It depends where exactly the cost sits and whether it’s backed by:
- model/data flywheels
- mechanical IP
- tight subsystem integration
In commoditized layers, cost savings often just pass through to the buyer.
But in leveraged layers, they unlock compounding advantage.
🧠 Takeaway: Cost only matters if it carries leverage
🧬 The invisible BOM matters more
What’s not in the BOM is often what defines long-term edge.
You won’t find fleet logs, ROS tuning configs, or simulation policies listed, but these layers determine:
- autonomy behavior
- adaptability in edge cases
- learning across robots
- margin expansion over time
This is the invisible layer:
data → model → behavior → more data
Whoever owns this loop doesn’t just reduce cost, they also bend the performance curve.
🧠 Takeaway: What’s not in the BOM might define long-term defensibility more than what is.
🗺️ How to read the BOM like a strategy map
BOM tells you exactly where leverage lives:
- Actuation → high cost, high moat, low margin
- Compute → low cost, high leverage, high margin
- Comms/frame → low cost, low moat, low margin
- Sensors → medium cost, increasingly commoditized
So the BOM becomes a proxy for margin profile:
- high BOM + high IP → hardware moat (gears, drives)
- low BOM + model/data ownership → software moat (GR00T stack)
- high BOM + low IP → worst of both worlds (commodity assemblers)
As robotics systems will evolve, I feel BOM will offer a uniquely grounded lens to assess technical constraints, economic structure, and strategic leverage of the robotics stack.
Thanks to @chynaqqq & @castorhat (PrismaX), @karsenthil, (Reborn), @xmercury_one (Xmaquina DAO), @ivailoj (Paper Ventures), @BlueHors3Shoe (No Limit Holdings), @shutterbugsid (Decentralised Co) for some quick feedback & suggestions on the piece.
4
15
749
Just joined the @adinlone Member network.
AI does the diligence.
Members surface the deals and earn carry.
Apply here → adin.online/waitlist
3
11
2,522






















